Home

A Test Setup for Comparison of People Flow Sensors

1. Clas Ohlson Logger 1 ae M Vh etwork Server Counting pulses In to cafe Out from caf Both directions Ethernet Figure 52 Block diagram of the sensor test s data collection arrangement 71 pame SS LEO Clients Web Browser Backup Po rs Pn 04 Clients Web Browser to Internet Figure 53 Operating principle of the Visit visitor counting data transfer system Mel ville amp Ruohonen 2004 Figure 54 Indicators for cumulative counting monitoring For direction sensitive sen sors the upper reading represents the to cafe and the lower the to corridor direc tion 12 5 4 Sensor configuration After the physical installations and mountings the sensors were configured appropriate ly In the case of the video and the IR camera sensors the settings were adjusted using individual software and appropriate connection cables After being properly assembled neither the pressure sensitive sensor mats nor the Clas Ohlson single beam IR sensor required any extra adjustment The beams of the Takex and Cedes IR sensors were adjusted to cover the test site width by trimming the sensor potentiometers The Takex sensor s illumination conditions switch was set in the Light ON position and the delay between adjacent counting pulses was selected as 0 2 s The Irisys thermal IR camera was configured using the Irisys IWC3052 setup module between the sensor
2. 3 Aa ee N SHR N N NN x AT EANN a NW W WW ES aS eee eam SE ONAL eee 4 NN dees JN NN Reflector SON TEA W Iik Photoelectric Max 6m Beam Sensor Figure 31 Clas Ohlson IR photocell sensor left and its plugging chart right Yu Heng Electric Co Ltd 2005 The Takex sensor used in the test setup was a combination of two Takex DL S202 background suppression IR photo sensor modules and a custom circuit assembled by Teknovisio Ltd Figure 32 left The circuit analyses the order of the pulses from the two sensor modules enabling direction sensitivity The device also contains an adjusta ble counting delay to reduce possible overcounting caused e g by swinging arms of passing pedestrians In practice the delay is created by adjusting the length of the out put counting pulse The sensor recognizes the movement direction from the interruption order of the two beams and sends a pulse to the corresponding output channel The sep aration of people walking side by side is however impossible The sensor uses a dis tance limited reflection method and hence needs no separate reflector plate The opera tion of a traditional beam sensor using a reflector plate is based on the received light intensity but in this case a two division photodiode is utilized to sense reflection dis tances from the angle of the received light Figure 32 right Takenaka Electronic In dustrial Co Ltd 2006 The detection range of both beams can
3. Bauer et al 2011 40 3 Applications for people flow sensors 3 1 Visitor counting and retail applications The shopper circulation ultimately determines the success of a shopping center and high customer flows are a prerequisite for a single stores success in a center Brown 1991 Des Rosiers et al 2005 Thus the knowledge of consumer movements is one of the central issues in shopping center business and the manipulation of customer flows has long been characteristic to shopping center planning One of the most essential rules is to place a mall s strategic elements like anchor tenants entry points and escalators in a way that contributes in the best possible way to the circulation of shoppers within the complex and hence maximizing the shoppers exposure to the offerings of different stores Brown 1991 One of the major findings in the retail industry has been that the placement of two department stores anchor tenants in one shopping center increased the businesses of both stores The value of information of people flows in retail is also supported by the fact that planning well a mall s tenant mix and both the individual and mutual placement of the tenants can further affect positively the sales of smaller non anchor tenants as well This is due to the fact that shops can generate sales from other stores customer traffic Eppli amp Benjamin 1994 Clustering similar stores increases their total sales and further the success
4. outdoor air entering the building Particulate pollutants like fibers can be released into indoor air by mechanical abrasion or air motion and by sneezing and coughing At equi librium indoor pollutant concentrations depend on their outdoor concentration indoor pollutant strengths and the rate of pollutant removal by ventilation Hence other types of pollutant sensors than CO sensors may be needed to maintain proper IAQ In a room where the occupancy level is variable but during occupied periods the number of peo ple is relatively stable occupancy sensors might be the most cost effective solution for a DCV with fixed low and high ventilation values Fisk amp De Almeida 1997 Tradi tional fixed ventilation however has been found to control formaldehyde concentra tions better than any of the DCV strategies As the most non occupant pollutant source strengths and limits are seldom known accurately an all time base ventilation rate or a morning purge have been recommended to complete selected DCV strategies To avoid the DCV system s reaction to non uniform pollutant distribution patterns it should also have a large time constant Emmerich amp Persily 1997 According to the ASHRAE American Society of Heating Refrigerating and Air Conditioning Engineers Standard 62 1989R the ventilation rate for an acceptable AQ is determined by the area of the ventilation controlled space and the estimated number of people in it As ventilation
5. Mysen et al 2005 The energy consumption measurements took into account the fan energy the heating energy due to ventilation heat loss and the ventilation heat energy recovery It was found out that during a ten hour operation period the CO sensor based DCV could re duce the energy use to 38 and the IR occupancy sensor based DCV to 51 of the fixed volume ventilation In reality the classrooms levels of use were usually 30 50 of the basic 10 11 hour ventilation s operation period In the case of six hour classroom use the energy savings for CO2 and IR DCV were 50 and 73 respectively and in the case of 24 hour use 26 and 32 respectively Thus the energy saving potential of a DCV system was sensitive to the fixed rate ventilation operation period An advantage of the CO sensor based DCV was the fact that it could reduce the risk of under ventila 47 tion in over crowded classrooms The IR occupancy sensors in turn were of a simpler technology and hence notably cheaper more robust and longer lasting Mysen et al 2005 Katabira et al 2008 proposed a system capable of providing cool or warm air supply only in places where it is necessary or where the people density is the highest It could be exploited in public spaces with continuous population density changes like railway stations or exhibition halls The air conditioning control system involved a wireless network of laser scanners and temperature sensors By integ
6. Sensor based demand controlled ventilation a review Energy and Buildings Vol 29 1 pp 35 45 ISSN 0378 7788 Greene Roesel R amp Di genes M C amp Ragland D R amp Lindau L A 2008 Effective ness of a Commercially Available Automated Pedestrian Counting Device in Urban Environments Comparison with Manual Counts In Transportation Research Board 87th Annual Meeting Washington D C USA 13 17 1 2008 Washington D C USA Transportation Research Board pp 1 15 Hashimoto K amp Yoshinomoto M amp Matsueda S amp Morinaka K amp Yoshiike N 1997 Development of people counting system with human information sensor using multi element pyroelectric infrared array detector Sensors and Actuators A Physical Vol 58 2 pp 165 171 ISSN 0924 4247 86 Hashimoto K amp Kawaguchi C amp Matsueda S amp Morinaka K amp Yoshiike N 1998 People counting system using multisensing application Sensors and Actuators A Phys ical Vol 66 1 3 pp 50 55 ISSN 0924 4247 Hashimoto K amp Tsuruta T amp Nishimura K amp Morinaka K amp Yoshiike N 1999 Development of One Dimensional Pyroelectric Infrared Array Detector with High Sen sitivity Japanese Journal of Applied Physics Vol 38 10 pp 5964 5968 ISSN 0021 4922 printed ISSN 1347 4065 electronic Hashimoto K amp Tsuruta T amp Morinaka K amp Yoshiike N 2000 High performance human information sensor Sensors and Actuator
7. USA University of California at Berke ley Institute of Transportation Studies California PATH 18 p Chen O amp Gao M amp Ma J amp Zhang D amp Ni L M amp Liu Y 2008 MOCUS mov ing object counting using ultrasonic sensor networks International Journal of Sensor Networks Vol 3 1 pp 55 65 ISSN 1748 1279 printed ISSN 1748 1287 electronic Chen T H amp Chen T Y amp Chen Z X 2006 An Intelligent People Flow Counting Method for Passing Through a Gate In 2006 IEEE International Conference on Robot ics Automation and Mechatronics Bangkok Thailand 7 9 6 2006 Piscataway NJ USA The Institute of Electrical and Electronics Engineers Inc pp 1 6 ISBN 1 4244 0025 2 Cognimatics AB 2010a Manual TrueView People Counter embedded for Axis IP Cameras Lund Sweden 28 p Referred 18 5 2011 Available http face cognimatics com downloads axis people counter manualT VPCAxisCam pdf 85 Cognimatics AB 2010b Camera based People Counter Brochure Lund Sweden 2 p Referred 14 10 2011 Available http www cognimatics com downloads trueview peo ple counter product sheet pdf Des Rosiers F amp Th riault M amp M n trier L 2005 Spatial Versus Non Spatial De terminants of Shopping Center Rents Modeling Location and Neighborhood Related Factors Journal of Real Estate Research Vol 27 3 pp 293 310 ISSN 0896 5803 Dharmaraju R amp Noyce D A amp Lehman J D 2001 An
8. algorithms also have been found to have good properties when there is background clut ter present in the handled image Frequent updating of the back image is needed for re liable motion detection The motion detection used by Heikkil amp Silv n 2004 was based on a pixel based absolute difference between the incoming and the adaptive background frames The pixels were assumed to contain motion if the absolute differ ence exceeded a certain threshold A Kalman filter based tracking method was used to determine the object correspondences between the consecutive frames The collected raw data also contained spatiotemporal coordinates that could be further utilized even in analysis of traffic accidents It was noticed that for an arbitrary installation of the track ing unit a general object classifier could not be created as the object shapes depended on the direction and the distance of the camera and road Instead to train the tracking software a learning vector guantization LVO method based on a user supervised learning principle was used Heikkil amp Silven 2004 For the field tests the tracking device was installed on a 3 m high pole about 10 m from the monitored road Figure 10 The achieved continuous counting time with a battery was 2 3 days The tests made in various weather conditions with simultaneous manual counting that lasted 0 5 3 h Both cases a path only for pedestrians and cyclists test 1 and a road side where mo
9. also significantly affects the energy consumption should over ventilation however avoided Leephakpreeda et al 2001 completed tests for an occupancy based DCV system s performance evaluation in two classrooms having ca pacities of 55 and 30 students The larger classroom was used to verify mathematical models for occupancy estimations Based on the activities of the people in the room the CO generation rate per person is usually known in advance and the number of occu pants in the room N can be calculated as follows Q c cs N G 6 where Q is the supply air airflow rate c and c are the indoor and outdoor CO2 concen trations respectively and G is the CO generation rate per person The use of this for mula however requires that the room s CO concentration reaches a sufficiently steady state If the activities change the generation rate must be adjusted accordingly In the smaller test room a real time ventilation control using the mathematical model men tioned above was tested Leephakpreeda et al 2001 For the testing of the mathematical models the CO2 concentration of the room was measured in five minute intervals The fresh air intake was kept constant at 288 s and the outdoor the CO concentration was measured to be 230 ppm In the beginning the ventilation was off and the CO2 concentration of the room was about 500 ppmy The value of the CO generation rate per person was 0 3 l min that is one for
10. and computer The installation height was set to 267 cm being the distance from the ceiling to the floor and thus egualing the dimension presented in Fig ure 46 plus the sensor s own height The discrimination sensitivity was set to a medium value 43 The large target couple counting was set to enable judging of larger than expected and right shaped objects as a pair of pedestrians walking close to each other The immediate count mode with the anti dither mode was used With this setting a count is registered immediately after a corresponding counting line is crossed instead of waiting until a person leaves the camera s field of view The configuration program and the counting line settings of the Irisys sensor are presented in Figure 55 Line 1 31 Line 2 39 aah Counts A ll Satine 4 Qirisys L 2com Countiine Actior Red Hot Intelligence Figure 55 Configuration software and counting line settings of the Irisys thermal IR camera sensor The counting parameters of the Sick time of flight camera were set as follows maxi mum and minimum person heights were 2 200 mm and 700 mm respectively and max imum and minimum person widths were 400 mm and 100 mm respectively The bor ders of the counting area were set to 1 100 mm Ymin and 1 265 mm Y max from the 73 sensor s center The sensor had no settings for the installation height The Sick sensor s configuration software displaying the sensor s environmen
11. applications Yoshiike et al 1999 The sensor was based on an eight electrode pyroelectric IR array detector fabricated from PbT103 ceramics It had a time constant of about 8 6 ms that was shorter than with commercially available pyroelectric detectors The measurement area of the detector was determined by the electrode size and the IR lens curvature in front of the detector The used electrodes were of different sizes the smallest corresponding to long distances and the largest to near locations in the monitored area The sensor head was located in side an umbrella shaped motor eguipped chopper Figure 25 left The chopper had three windows at 60 degree angles to each other Thus it allowed three open close choppings during one full rotation The sensor was able to create a two dimensional thermal image of the entire span of the room with a resolution of 8 x 120 pixels Linear rotary scanning and a fuzzy clustering algorithm were used to identify the number of people in the image The sensor was capable detecting objects both at lower or higher temperature than the ambient condition The view range of the sensor can be seen on the right in Figure 25 Yoshiike et al 1999 2 7m sensor position 1 0m 0 0m umbrella shaped chopper Figure 25 Structure left and view range right of an eight electrode pyroelectric IR detector by Yoshiike et al Yoshiike et al 1999 5m 4m 3m 2m 1m Om 46 The sensor was teste
12. centage error was reduced to 3 with up to ten people and to 1 5 with up to eight people Amin et al 2008 Figure 12 People counting setup involving a video camera and an IR camera by Amin et al Amin et al 2008 2 7 Ultrasound sensors Ultrasound sensors are capable of working in both indoor and outdoor conditions re gardless the ambient temperature or lighting conditions Chen et al 2008 developed a direction sensitive system called Moving Object Counting approach using Ultrasound Sensor networks MOCUS employing three node ultrasound sensor clusters Each of 23 the clusters had one transmitter and two receiver nodes connected to a wireless MICA2 multidimensional independent component analysis measurement unit While one mov ing object usually caused multiple detection events of one reflected ultrasound pulse a certain threshold time was used to classify the detection events originating from sepa rate pedestrians The direction of object movement could be solved with two ultrasound nodes placed successively and parallel to the monitored thoroughfare and by comparing their detection timestamps Too closely placed multiple ultrasound transmitters howev er could cause unpredictable interference with each other Figure 13 a A solution to this problem was introduced in a three node sensor cluster where one node acted as a transmitter and the other two were receivers Figure 13 b Receivers also received less noise w
13. evaluate a case study algorithm a stereo based preprocessing for human shape localization The rate of cor rect detections was found to be about 83 with around 0 46 false positives per frame It was mentioned that higher thresholds in the algorithm would have decreased the num ber of false positives but reduced the correct detection rate as well The main results of the tests were statistical characterization of the detection algorithm and identification of situations that were challenging to it The manual annotation could be further improved by introducing an automatic preliminary bounding box placement Bertozzi et al 2004 N f NA ih al a N i f N gt Video sequence specific algorithm performance Annotation lt gt FileH Annotation File A Statistical Analysis MN i Video Sequence Annotation Engine Algorithm i Annotation Tool a Figure 15 Block diagram of a calibration procedure for the vision based pedestrian detection system by Bertozzi et al Bertozzi et al 2004 Ruph Haller et al 2006 installed pressure sensitive sensor slabs at the four main en trance locations of the Swiss National Park in Graub nden Switzerland A four day span of sensor data was collected simultaneously with manual control counting by hu man observers The sensors were buried under 8 10 cm of soil and connected to data loggers collecting counting data having one hour
14. evaluation of technologies for automated detection and classification of pedestrians and bicycles In The 71st ITE An nual Meeting Chicago IL USA 17 21 8 2001 pp 1 13 Efmit Ltd 2003 Emfit L series sensors specifications Rev C Datasheet Vaajakoski Finland 2 p Referred 18 5 2011 Available http www emfit com uploads pdf Emfit_ L series specifications pdf Emmerich S J amp Persily A K 1997 Literature Review on CO Based Demand Controlled Ventilation In The 1997 annual meeting of the American Society of Heat ing Refrigerating and Air Conditioning Engineers Boston MA USA 28 6 2 7 1997 Atlanta GA USA The American Society of Heating Refrigerating and Air Conditioning Engineers ASHRAE Transactions 1997 Vol 103 2 pp 229 243 ISBN 978 9 9962 2378 5 Eppli M amp Benjamin J D 1994 The Evolution of Shopping Center Research A Re view and Analysis Journal of Real Estate Research Vol 9 1 pp 5 32 ISSN 0896 5803 Eurotech S p A 2010 PCN 1001 Passenger amp People Counter User Manual Rev 3 1 Amaro Italy 98 p Referred 18 5 2011 Available http www eurotech com en pb aspx tab download amp mm devices amp tmm People 20Passenger 20Counters amp pg P CN 1001 Eurotech S p A 2011 PCN 1001 Passenger Counter Datasheet Amaro Italy 2 p Referred 17 10 2011 Available http www eurotech com DLA datasheets Products_ Eurotech PCN 1001 sf pdf Fisk W J amp De Almeida A T 1998
15. general used to calibrate automatic counting sensors Cessford et al 2002 2 2 Infrared light beam sensors Song et al 2008 examined a customer counting system based on a radial basis func tion RBF neural network using IR photoelectric sensors Video counting has to handle a large set of image data and a complex counting algorithm Hence it has difficulties in satisfying real time rate and accuracy requirements Strict environmental requirements for lighting conditions also limit the usefulness of the method Active IR counting tech nology is considered to be more stable as it requires less handling of measurement data and is resistant to environmental effects IR beam sensors however cannot distinguish people entering the counting location simultaneously Song et al 2008 To improve the counting accuracy of IR beam sensors Song et al presented a customer counting system based on an RBF neural network They assembled four IR beam sen sors on the two sides of a market s doorway The sensors were placed at the height of an ankle and at a distance of 25 cm from each other Figure 2 above The reaction time of the sensor type used was 30 ms For the collected data a noise elimination and a stand ardization procedure was used The pulse widths and intervals along with the mutual order of the four sensors were used as a training input to the RBF neural network The neural network was taught the signal patterns caused by different numb
16. manual real time counting results A sample shot of the video capture is shown in Figure 58 in section 5 4 The control counting from the video capture has certain advantages over the real time manual counting In unclear cases the video re cording can be reviewed to confirm the number of visitors or even paused to take a break in the counting The latter helps to reduce the numbing effect of the task and pre vents possible loss of concentration of the monitoring persons The video is not availa ble from the morning counting period of Friday 8 4 2011 due to a user mistake when preparing the computer for the video capture Figure 59 Four digit hand held counter 76 Table 5 Periods of manual control counting Day Morning Afternoon Wednesday 30 3 2011 9 00 10 00 12 00 13 00 Friday 8 4 2011 10 00 11 00 13 00 14 00 Monday 11 4 2011 14 00 15 00 6 2 Counting reports The collected sensor test data was available in one hour sampling intervals on the Visit Service web portal automatically arranged under a corresponding sensor After logging onto the web site all test sensors were listed in a table Figure 60 By clicking the name of a particular sensor its counting results were displayed Figure 61 If the sensor was direction sensitive both the to caf and the to corridor directions appeared and total readings were automatically displayed separately The resolution of the counting data could be sel
17. no power supply as such but in this test setup it was sup plied with a voltage to create a counter pulse when stepped on K A Schmersal GmbH 2010a 2010b E n Contact Figure 30 Schmersal switching device mat sensor left and its operating principle right K A Schmersal GmbH 2010b 55 4 3 Infrared light beam sensors Clas Ohlson DES 700 is an IR photocell sensor without direction sensitivity Figure 31 left The sensor needs a reflector on the opposite side of the monitored corridor Its plugging chart is shown on the right in Figure 31 The Clas Ohlson sensor comes with a power supply and a compact central unit containing an adjustable electronic chime and speaker This makes possible utilizing the device as an automatic doorbell for retail ap plications or as an alarm for some simple security solutions The central unit can be eguipped with a separately sold counter with a four digit indicator A passer by blocks the emitted IR beam s reflection back to the sensor hence causing a counter pulse For the automatic data collection needed in this test setup the sensor s cabling was modi fied to make a connection to the pulse wires Yu Heng Electric Co Ltd 2005 adapter DC9V M EISIEISIRIN II 4 ja 2 AALS Ke 5 coe Hais arli p Vr ean CEEI Doar 4 TO Hi OUTPUT 7 Twin Plugs Shelter Speaker cable wire j W ERI m NN gt x n gt 3 3 5 W NN SaR
18. of counted people in the case of the IR beam sensor was due to the sensor mis placement mentioned before When aggregating the passenger flow to 15 minute inter vals the residual standard deviation of the IR beam s and switching mat s reading error was 4 6 persons and 5 6 persons respectively at the security checkpoint and 4 5 per sons and 3 4 persons respectively at the border checkpoint At the border checkpoint however the count was unreliable due to the small sample size Applying a case specific calibration factor to the sensors readings they would be able provide reasona ble accuracy for aggregation of several hours data even if the counter flow was taken into account For short term applications like online crown control systems the count ing errors were nevertheless too large Bauer et al 2011 The service time analysis showed that one directional counting led to substantial service time underestimation The situation was worse with the switching mat The interperson al spacing was seen to vary with the number of persons served in hurried situations A calibration factor estimated to equate the mean of the interpersonal spacing was intro duced to proportionally increase the interpersonal spacing from the sensors and to com pensate the incorrect crossings If the derived interpersonal spacing distributions of the sensors were used in checkpoint capacity simulations the differences with manual dis tributions were minor
19. peak hours to get a sufficient amount of test data Tests included two bicycles with ferrous metal chrome alloy and non ferrous alu minum Al bodies Observations included 100 one way walks 100 ferrous metal and 51 non ferrous metal bicycle trips collected in three separate periods SRF Consulting Group Inc 2003 The sensor counting results were compared against ground truth manual counting data collected by multiple observers For pedestrian detection the counting errors were 0 for the Autoscope Solo the MS Sedco and the ASIM sensors and 7 for the Diamond sensor Ferrous metal bicycle detection counting errors were 1 for the Autoscope Solo 4 for the MS Sedco 1 for the ASIM 4 for the Diamond 0 for the In ductive Loop Detector and 2 for the 3M Microloop sensors The 3M Microloop s second lane probes were out of order so its sample included data of only 50 ferrous metal bicycles Non ferrous bicycle detection counting errors were as follows Auto scope Solo 0 MS Sedco 2 ASIM 0 and Inductive Loop Detector 0 The 3M Microloop sensor was left out of the non ferrous detection test as it could only detect ferrous metal objects The exclusion of the Diamond sensor was not explained The Diamond sensor however has a wide operation range and can thus be installed tens of meters away from the monitored trail It was concluded that for real world tests a more versatile selection of detected objects and a longer monitoring perio
20. people The sen sors were assembled as previously above the monitored doorway at intervals of 60 cm Altogether 1 593 persons walked freely through the doorway during an observation pe riod of six hours The sensor s chopping frequency was kept the same as in the case of one sensor 20 Hz The achieved minimum detectable distance between two persons was about 10 cm and the detection accuracy of movement direction and of the number of passers by was 99 and 95 respectively It was also concluded that thinning of the film thickness of pyroelectric material could improve its sensitivity Also using a new algorithm taught with the thermal images of new passing objects would increase the counting accuracy Hashimoto et al 1998 Quantum type IR sensors have high sensitivity and resolution although to reduce the thermal noise they need to be cooled down during operation They are also expensive in cost and quite impractical due to their large physical dimensions Uncooled thermal type IR sensors like thermopiles resistive or dielectric bolometers and pyroelectric sensors are still very suitable for human detection In their research Hashimoto et al 12 1999 presented an IR sensor that has 16 PbTiO elements double in number to their previous one The high accuracy of the sensor was achieved by a precise sheet forming method that could keep the sensitivity variations between different sensor elements at less than 10 The characteristi
21. the front panel of the device must be placed parallel to the floor The device can be recessed even on non horizontal surfaces and tilted 0 20 degrees up to 45 degrees using an extender A maximum of six PCN 1001 sensors can be connected together to cover corridors and entrances wider than 120 cm For every extra 60 cm an extra device is needed The detection area of the sensor 1s 100 cm in height and the distance between the front panel and the detection area s upper limit can be selected between 25 and 40 cm Figure 36 right The height of the people to be counted thus has to be taken into account when installing the device The sensor can be configured using individual client software connecting a computer to the Mini USB 1 1 port located under a service plate on the sensor s front panel Figure 36 left The special visual tracking mode of the software can be used to verify the proper operation of the sensor Eurotech S p A 2010 FRONT PANEL PERPENDICULAR AXIS FRONT PANEL LONGITUDINAL AXIS FRONT PANEL TRANSVERSAL AXIS Parallel to the door i Perpendicular to the door f 90 90 si 25 30 to 31 40 cm Enclosure Infrared LED emitters Service plate Frame 100 cm Front panel with adjustable angle Stereoscopic cameras sensors N 2H tg 70 2 322 cm when H 230 cm D 2H tg 55 2 le 240 cm when H 230 cm E i J N FLOOR DETECTION AREA without masking Figure 36 Front panel of the E
22. the scenario with a camera installed over a gate a random number up to five test subjects fully passed through the camera s field of view Altogether 30 tests were conducted five times with every pedestrian set combination Special cases with one person standing at the monitored scene and a per son stopping at the scene and returning back were also used The other two tests were static and dynamic versions of elevator scenarios In the static case the test subjects up to five people stood in an elevator and in the dynamic one they up to ten people were also ordered to enter or exit the elevator cabin Plain background images of the scene without any test subjects were taken at each scenario as reference images The training of the IR ANN was done by using twenty samples and the performance tests of the sys tem were completed using 200 samples from every experiment mentioned above The visual ANN was trained using background images only Amin et al 2008 The detection error of the IR ANN with up to ten people was within 4 5 and of the visual ANN with up to six people within 5 When more people were present they stood very close to each other causing the visual ANN a detection error of around 12 The IR system was found to be more capable of handling higher people densities with higher accuracy and the visual system was more reliable in the lower density cases Combining the results of both the IR and visual counting system the maximum per
23. the start and the end of the ser vice of a single passenger Bauer et al 2011 Both the sensors were tested simultaneously and connected to a PC through separate measurement units containing a nonlinear delay circuit and an analog to digital con verter Demo software was used to register the sensor pulses in real time and to save them with time stamps At the security checkpoint placement options of the sensors were heavily restricted The IR beam sensor was mounted next to the metal detector 38 after the conveyor belt L3 in Figure 22 left As the switching mat could not be placed in the metal detector it was assembled next to the conveyer belt in an area where pas sengers typically take off their coats and metal containing objects the green rectangle in Figure 22 left At the border checkpoint the IR beam sensor was mounted on fences delimiting the walking lines Figure 22 right and the switching mat was placed in front of the checkpoint booth the green rectangle in Figure 22 right At both checkpoints sensor counting data was collected for about three weeks However partly due to the failure of the sensor mat s power supply and misplacement of the beam sensor only about 7 5 days of data was gathered from the border checkpoint Bauer et al 2011 Figure 21 IR light beam sensor left and switching mat sensor right tested by Bauer et al Bauer et al 2011 Figure 22 Sensor test sites by Bauer et al security
24. to compare pedestrian counts registered at different locations timeslots and under different weather conditions Prevailing gasoline prices were actu ally collected during the tests as it could have an effect to the pedestrian volumes The tests should also be repeated at a greater variety of locations and in different environ mental conditions to estimate the error margins of the extrapolation procedure Reliable weekly estimates of pedestrian volume could be exploited by authorities to improve the safety and convenience of pedestrian traffic Schneider et al 2009 2 5 Pressure sensitive sensors In visitor counting methods managers of conservation areas and national parks desire reliability portability light weight accuracy robustness easy concealment water resis tivity temperature insensitivity minimal mechanics and electronics low maintenance needs low cost and low power consumption Sophisticated data e g direction of travel or timestamps are not their top priority For this purpose Cessford et al 2002 devel oped a step counter a pressure sensor built in the front board of a soil or wooden step or multi step structure Figure 8 Cessford et al 2002 STEP UNIT i Rear View in a step sequence install Step Unit here Sensor Ampi Hier a Cable with plug Visitor Counter Data download cable to VAMS datalogger Figure 8 Step counter construction left and sensor placement right by
25. videotaped the whole test and the video recordings were later analyzed by researchers All passing people including pos sible bicycles and infants in prams were counted Greene Roesel et al 2008 Counter Imaginary Line Figure 6 One of the three test sites by Greene Roesel et al Greene Roesel et al 2008 The videotapes were reviewed many times so theses counting numbers were assumed to correspond to the ground truth Judging when a pedestrian crossed the virtual count ing line however was occasionally difficult as some people wandered around the sen sor e g while using a cell phone In these cases the sensor was also sometimes ob structed and unable to see other pedestrians Some video recording periods were miss ing due to tape change and obstructed camera so the average pedestrian volumes for the sites were counted from manual field counting For sites 1 2 and 3 these were 2 614 223 and 1 464 respectively Test sensor readings and manual field counting volumes were compared against the video counting results according to Equation 3 The counting error for sensors and manual counting was calculated for each 15 minute period and for the total four hour periods Greene Roesel et al 2008 Manual field counting errors were between 1 4 0 4 These results differed signifi cantly from the inter reliability results of manual field and video recording counting by Diogenes et al 2007 that were between 15 8 In
26. 00 00408CAD3D7F Axis 00408CAD3D7F 8 10 20110103200000 00408CAD3D7F Axis 00408CAD3D7F 5 8 20110103210000 00408CAD3D7F Axis 00408CAD3D7F 3 2 20110103220000 00408CAD3D7F Axis 00408CAD3D7F 0 0 20110103230000 00408CAD3D7F Axis 00408CAD3D7F 0 1 Figure 62 Example of the one hour resolution data file structure of the Cognimatics counting software embedded in the Axis network camera sensor The first column is the time stamp and the last two columns are the values of the in and the out counting directions respectively 78 If both counting directions and their combined values were accounted as separate read ings datasets with 864 and 2 592 hourly readings in a one hour resolution were gath ered for the direction insensitive and direction sensitive sensors respectively For the nine test sensors this makes 18 144 hourly sensor readings in all 6 3 Precision test data Although the highest resolution sensor test data available on the Visit server was one hour six minute resolution data was separately received from Teknovisio Ltd This pre cision data is logger specific that is every logger has its own report file containing read ings of the sensors connected to it from the entire test period An example of the preci sion data is presented in Figure 63 This data can be used for more specific sensor per formance analysis and to identify possible inconsistencies between the readings of dif ferent sensors and any one sensor itself N
27. 150 2 100 The Clas Ohlson IR beam sensor was purchased from a re tail store and had a price of about EUR 50 including taxes The sensor prices are com pared in Figure 64 No mutual comparison inside the low and high capacity catego ries is at this point done While examining the prices it should be kept in mind that the Clas Ohlson sensor and the mat sensors are not capable of sensing the movement direc tion High Capacity 0 500 1000 1500 2000 Price EUR Figure 64 Comparison of the test sensors wholesale prices For the Clas Ohlson sen sor a retail price with taxes is given No mutual comparison inside the high and low capacity categories is done Assembly of the test sensors in an indoor location does not set any special requirements if suitable floor wall or ceiling surface 1s available However notice of proper and se cure mounting should be taken especially in case of the heavy weighted overhead cam era sensors The test equipment including the sensors data logging units and cabling should also be properly shielded against the mechanical abrasion and possible vandal ism The Emfit sensor mat is very thin and can be installed under a floor material or a rug but if the thicker Schmersal mat is used an additional ramp rails might be needed to avoid general mechanical stress and possible tripping of pedestrians 82 7 2 Future work Future work will include the analysis of the people flo
28. 2011 Available http www axis com files manuals um M32series 39122 en 1004 pdf Bauer D amp Br ndle N amp Seer S amp Ray M amp Kitazawa K 2009 Measurement of Pedestrian Movements A Comparative Study on Various Existing Systems In Tim mermans H ed Pedestrian Behavior Bingley United Kingdom Emerald Group Publishing Ltd pp 325 344 ISBN 978 1 84855 750 5 Bauer D amp Ray M amp Seer S 2011 Using Simple Sensors for Measuring Service Times and Counting Pedestrians Strengths and Weaknesses In Transportation Re search Board 90th Annual Meeting Washington D C USA 23 27 1 2011 Washing ton D C USA Transportation Research Board pp 1 13 Bertozzi M amp Broggi A amp Grisleri P amp Tibaldi A amp Del Rose M 2004 A tool for vision based pedestrian detection performance evaluation In 2004 IEEE Intelligent Vehicles Symposium Parma Italy 14 17 6 2004 Piscataway NJ USA The Institute of Electrical and Electronics Engineers Inc pp 784 789 ISBN 0 7803 8310 9 84 Brocklehurst D 2005 People Flow Modelling Benefits and Applications within In dustry Doctoral dissertation Loughborough United Kingdom Loughborough Univer sity Centre for Innovative and Collaborative Engineering 136 p Brown S 1991 Shopper Circulation in a Planned Shopping Centre International Jour nal of Retail amp Distribution Management Vol 19 1 pp 17 24 ISSN 0959 0552 Bu
29. 2011 The shortest available six minute inter val sampling rate was collected No manipulated patterns of passers by were used but people were freely allowed to walk through the test site as usual The collected test data will be used to compare the people flow sensors mutual accuracy and overall counting performance The analysis of the test data however was left as a future research topic In addition to the automatic sensor data collection a manual on site control counting was carried out in five one hour periods Two volunteers were seated in suitable loca tions on both sides of the test site doorway each monitoring and counting only one of the two possible walking directions through the doorway By doing this any problems involved in using one person to monitor two directions simultaneously were avoided The manual monitors used simple four digit hand held counters to register the passers by Figure 59 The manual counting took place during three days as presented in Table 5 The manual results will be used as ground truth data in the future analysis of the col lected sensor test data Simultaneously with the manual counting a video recording of the test site was cap tured For this purpose the live view function of the Cognimatics TrueView software embedded in the Axis network camera sensor was utilized Video capture from the test site was recorded on the Ethernet connected PC The video footage can be used later to double check the
30. 247 Yu Heng Electric Co Ltd 2005 DES 700 Door Entry System Alarm Datasheet Kaohsiung City Republic of China Taiwan 2 p Referred 18 5 2011 Available http www e scs com tw eng product pic_04 catalog DES 700 pdf Yu Heng Electric Co Ltd 2008 Manual of Door Entry Alarm amp Counting System Item No DES 700 DC 500 Kaohsiung City Republic of China Taiwan 2 p Re ferred 18 5 2011 Available http www e scs com tw eng techonlogy download DES 700_DC 500 pdf
31. 63190002 2 Takex kahvilasta 4 4 2011 9 12 4 4 2011 9 18 2 OPEN VI3004604T63190002 3Cedec kahvilaan 4 4 2011 9 12 4 4 2011 9 18 1 OPEN VI13004604T63190002 4 Cedec kahvilasta 4 4 2011 9 12 4 4 2011 9 18 2 OPEN VI13004604T63190002 1 Takex kahvilaan 4 4 2011 9 18 4 4 2011 9 24 9 OPEN VI13004604T63190002 2 Takex kahvilasta 4 4 2011 9 18 4 4 2011 9 24 6 OPEN V13004604T63190002 3 Cedec kahvilaan 4 4 2011 9 18 4 4 2011 9 24 9 OPEN VI13004604T63190002 4 Cedec kahvilasta 4 4 2011 9 18 4 4 2011 9 24 7 OPEN Figure 63 Example of a six minute resolution test data For every sample period con nected logger channels A and their data values B are shown 79 7 Conclusions and future work 71 Conclusions Therange of people flow sensors include several technologies that differ in their charac teristics installation reguirements counting capacity and price No single sensor type is suitable for every possible location environment and application Sensor development and single sensor performance tests have been performed on a large scale as have re views of commercially available devices However these reviews rely mainly on the sensor specifications reported by the manufacturers or vendors The number of more comprehensive surveys including comparative tests of multiple sensors is more lim ited Some of these publications give only partial information about the sensors under tests e g sometimes only the sensor brand is given and the
32. 66 URSKA 1 20 E 8 96 Z 8 56 p O il g N br TT x Ic I i TT I L a I I B TI Tr T 136a TUTKIMUSHALLI i uae E J D EI60 P 6 J n HL 113 KH osan 2 FI v I NEUVOTTELUH a a x S manr J HE mi XE 111 HUOLTOLUUKKU E160 114 vw VARASTO f ji fo 13607 Cafe w 2 05 ss 34 Sa KENNE Entrance on h 2 20 1 32 SISAANTULOHALLI Wi 1 56 ws A PORRAS 28000 Caf 133c E J TOIMISTOH TY KALUT A 103 TOIMISTOH JUL vee TM Ci JI IF I 4 T G H 102 AE TOIMISTOH 105 H W TOIMISTOH B D 7850 L 20750 lj 104 5 6 Fi 8 9 10 TOIMISTOH all j l HATAPOISTUMINEN IKKUNASTA At L J O 7 25 5 93 N 12600 it 7 50 Entrance YL IKKUNAT gt LLEEIILIIITILIIETILIILIITVILLIITISLIIITIISTIEAITELTINITILTIIIEISIILIS ESTI TILL EN IILSLITSILILILIIIIIILLILTIISLILAITIIIENNTLINTIIELISLIITLLITALISTIEI ISI IIEISIEISILTIIISIIIEILIITELIITIITIILEITLIITEI TEITTE ISIIITEISSIISISON PAIK NIRS NRO SLO RIA SNAIL LAL GL AL SRL SL LS RGR AL SIAL RL OL LLL SRR OER LINGE z TN 152 S H 4 ISSN ISIST RN AN N H ET euman reusan Gg E Vi ren 1354 rw
33. AALTO UNIVERSITY School of Electrical Engineering Department of Electronics Jussi Kuutti A Test Setup for Comparison of People Flow Sensors Licentiate s Thesis submitted in partial fulfillment of the requirements for the degree of Licentiate of Science in Technology in Espoo 23 3 2012 Supervisor Professor Raimo Sepponen Second examiner Professor Petri Vuorimaa Instructor Lic Sc Tech Matti Linnavuo AALTO UNIVERSITY ABSTRACT OF THE SCHOOL OF ELECTRICAL ENGINEERING LICENTIATE S THESIS People flow sensor is a device used to measure the number and often also the direction of people pass ing through a certain passage per unit time Sensors based on a variety of technologies are available including infrared beam sensors and video and infrared camera sensors The applications of people flow sensors include gathering counting data for customer flow modeling demand based ventilation control and safety applications The quality and further feasibility of the collected sensor data are subject to the accuracy and reliability of the sensor used That is why the proper sensor for a desired application has to be chosen Efficient state of the art sensors are needed in busy locations but are usually expensive and often require stable environmental conditions Simpler budget priced sensors cannot handle complex situations but are generally more robust in tolerance to diverse ambient condi tions The aim of this thesis is to design and
34. Cessford et al Cessford et al 2002 The solution was based on video and field observations showing that almost none of the walkers missed the leading edge of the lowest step of the stairway This could be further ensured by increasing the stair height to 20 22 cm The counting unit was described as robust simple waterproof lacking moving parts and cheap less than EUR 300 It also 17 included storage capacity of thousands of records with timestamps Verified by field and video observations approximately 95 of descending and about 80 of ascending walkers stepped on the sensor giving an overall hit rate of 85 90 Counting errors were mentioned to be largely constant and hence they could be estimated by calibra tions Cessford et al 2002 Melville amp Ruohonen 2004 inspected an outdoor visitor counting system involving a vulcanized rubber shielded waterproof pressure sensitive Visit Mat sensor Figure 9 connected to a GSM based global system for mobile communications Visit Log data harvesting system and needing minimal maintenance work Pressure mats or pads can be concealed in the soil or pavement of the trail and are hence inconspicuous A draw back is a potential freezing over of the covering soil causing malfunctioning of the sen sor Melville amp Ruohonen 2004 The operation of the counting system was evaluated on twelve National Nature Re serves in different parts of England by English Nature a government age
35. Enclosure channel voltage Vpc consumption temperature rating status W C Emnfit L series Low 12 From logger 0 02 20 50 N A custom Schmersal SMS 4 Low 12 From logger N A 0 60 IP65 Clas Ohlson DES 700 High 9 15 AC adapter 0 7 max 20 50 N A Takex DL S 202 Low 12 From logger 0 36 max per 51455 IP66 custom sensor module Cedes 9 TPS 210 Low 10 30 From logger 0 4 max 40 65 IP65 Irisys IRC3020 Low 10 28 From logger 1 1 12 V 0 40 N A Eurotech PCN 1001 High 9 32 AC adapter 15 max 25 55 IP65 Sick TVS 100 High 24 15 AC adapter 69 8 typical 0 50 IP67 sensor with pulse IP40 fan outputs Axis M3203 N A 48 AC adapter N A 0 50 N A Table 4 Electrical and ambient properties of the test sensors 51 PoE Class 2 a Emfit Ltd 2003 Teknovisio Ltd 2009 b K A Schmersal GmbH 2010a 2010b c Yu Heng Electric Co Ltd 2005 2008 d Takenaka Electronic Industrial Co Ltd 2006 e Cedes AG 2009a 2009b f InfraRed Integrated Systems Ltd 2009a 2010 g Eurotech S p A 2010 2011 h Sick AG 2009 2011 1 Axis Communications AB 2008 2010a 2010b Cognimatics AB 2010a j Power over Ethernet k Sensor modules only 4 2 Pressure sensitive sensor mats The Emfit L series sensor mat consists of an elastic piezoelectric film placed between signal and ground electrode plates made of aluminum and covered on both sides with a shielding p
36. F amp Greene Roesel R amp Diogenes M C amp Ragland D R 2007 Estimating Pe destrian Accident Exposure Automated Pedestrian Counting Devices Report Research report Richmond CA USA California PATH Berkeley CA USA University of Cal ifornia Traffic Safety Center 28 p Cedes AG 2009a CEDES TPS Triangulation Proximity Switch Installation and opera tion manual Rev 1 1 Landguart Switzerland 12 p Referred 18 5 2011 Available http www cedes com pdfs produkte dg 106687en Operating manual TPS pdt Cedes AG 2009b TPS product Family Datasheet Landguart Switzerland 2 p Re ferred 14 10 2011 Available http www cedes com pdfs produkte dg 001133en TPS pdf Cessford G amp Cockburn S amp Douglas M 2002 Developing New Visitor Counters and their Applications for Management In Arnberger A amp Brandenburg C amp Muhar A eds Monitoring and Management of Visitor Flows in Recreational and Protected Areas Conference Proceedings Vienna Austria 30 1 2 2 2002 Vienna Austria Bo denkultur University Vienna Institute for Landscape Architecture and Landscape Man agement pp 14 20 Cessford G amp Muhar A 2003 Monitoring options for visitor numbers in national parks and natural areas Journal for Nature Conservation Vol 11 4 pp 240 250 ISSN 1617 1381 Chan C Y amp Bu F 2005 Literature Review of Pedestrian Detection Technologies and Sensor Survey Mid term report Berkeley CA
37. S merion tinon aan T A OEA 4 2 PREVIOUS RESEARCH ON PEOPLE FLOW SENSORS ssssoeeccssssssocooccccssssso 5 Zak SENSOR TY PES eiar i outdo apse mecenraseneenaat ices A E 5 2 2 INFRARED LIGHT BEAM SENSORS ammuttua ccesGuradeassatonstnavdaden tears 4 kv maa o Da a A 8 2 3 ACTIVE INFRARED SENSOR ARRAY Sorini ae raa aamun saber necaimaabicanes dues Samm kenk otava lamaa alaa 9 24 PASSIVE INFRARED SENSORS ioe sic asuteta ma Tassa aa ops word ao E AAG 10 29 PRESSURE SENSITIVE SENSORS Saj akvasmtouuaa hva kuu ETE OEA EE EET 16 2 6 VIDEO AND INFRARED CAMERA SENSORS ooooouuse nao aa aa aa naama aa naa a a aa aa a nanna naene 18 2l VETRASOUND SENSORS ur n E E E T O TT O NA 22 2 8 SENSOR CATIBRATIONS b ismomustusudim ousosuiasson E ao niUo E tonUaa O go iloa a NM NOT OSTAA sS TUON 25 2 9 SENSOR REVIEW Sroine skalvid ss k ali akt det mit valaialaussa mismaroikiev mesi yokaa en a k y Raket 2 2 10 SENSOR COMPARISON TESTS arm haaa impe N tama sulassa Oma Kuna uEt A 30 3 APPLICATIONS FOR PEOPLE FLOW SENSORS eeesssoscccsssoocccscssocceesesssoe 40 3 1 VISITOR COUNTING AND RETAIL APPLICATIONS ooouuoo omaa a aa aan a aa a ae aan a aa naan 40 3 2 DEMAND CONTROLLED VENTILATION uhiaossnta va aasuma laatatsuvaokaeoktae Sja va thansnteonelenage maaan series 43 3 3 SAFBTY APPIICATIONS acerra aE NOO saama vane Sameena km kotoa mP mat Saas ens 48 4 SENSORS OF THE TEST SETUP uaaosaoassvussasaaannuuaaaana annu unaaaa annu
38. SEDCO were based on microwave tech nology and from ASIM on passive IR technology Sensors from both manufacturers were mounted on a traffic light pole and tested at both curbsides and crossing areas of one pedestrian crossing Figure 18 Additional test hardware included a computer a pan tilt zoom video camera for live video capture and data processing equipment Data from all sensors were collected through an Ethernet connection and was available at one second resolution Detection zones for the sensors were identified based on engi neering judgment and specifications provided by the manufacturers Turner et al 2007 Figure 18 Curbside detection zone in red and crosswalk detection zone in blue used by Turner et al 2007 US Department of Transportation 2011 The output of the presence detection sensors is in most cases binary detection or no detection Thus three pedestrian counting sensors collecting a total people count were additionally selected for separate trail based test setups These included the Jamar 34 Scanner passive IR sensor the TrafX Infrared Trail Counter and the Diamond Traffic TTC 4420 pulse IR counter with a reflector The counting sensors were tested at two sites at College Station Texas and at one site at Austin Texas The sensors were mounted on available fixed objects like poles At the first site a controlled test was conducted by varying the target speed stopping to talk walking jogging running a
39. Schmersal C Clas Ohl son E Cedes F Irisys and G Sick 63 Figure 41 IR beam sensors as seen from the corridor C Clas Ohlson D Takex and E Cedes Figure 42 Camera sensors as seen from the floor The caf is located to the left and the corridor to the right Sensors F Irisys G Sick H Eurotech and J Axis 64 The two pressure sensitive sensor mats were assembled on the corridor side of the site and one on the other The Emfit piezoelectric mat was placed between two thin rubber rugs that were attached to each other and onto the floor with a strong double sided ad hesive tape The rubber rugs were used to shield the thin sensor mat that is convention ally installed under flooring material or encapsulated The Schmersal switching mat was fastened with tape on the top of the Emfit mat The placement of the sensor mats with essential measurements is shown in Figure 43 The reason why the sensor mats were placed some distance away the doorway threshold was because of the space needed by the rubber rugs shielding the piezoelectric mat The boundaries of the rubber rugs are left out from the sketch in Figure 43 for clarity Caf 170 cm 156 cm 1 5 cm N Threshold N 14 cm lt 120 cm 4 Corridor Figure 43 Placement of the pressure sensitive sensor mats view towards floor Sen sors A Emfit and B Schmersal The three IR beam sensors were mounted on the test site doorframe and adjacent wall as shown in Figur
40. a Research Centre Helsinki Research Unit pp 38 44 ISBN 951 40 1930 X Meyn S amp Surana A amp Lin Y amp Oggianu S M amp Narayanan S amp Frewen T A 2009 A Sensor Utility Network Method for Estimation of Occupancy Distribution in Buildings In Joint 48th IEEE Conference on Decision and Control CDC and 28th Chinese Control Conference Shanghai China 16 18 12 2009 Piscataway NJ USA The Institute of Electrical and Electronics Engineers Inc pp 1494 1500 ISBN 978 1 4244 3872 3 Mysen M amp Berntsen S amp Nafstad P amp Schild P G 2005 Occupancy density and benefits on demand controlled ventilation in Norwegian primary schools Energy and Buildings Vol 37 12 pp 1234 1240 ISSN 0378 7788 Newman A J amp Yu D K C amp Oulton D P 2002 New insights into retail space and format planning from customer tracking data Journal of Retailing and Consumer Ser vices Vol 9 5 pp 253 258 ISSN 0969 6989 Paavola L 2010 Pedestrian movement and modelling in a shopping centre environ ment Version 3 Literature review Espoo Finland Aalto University School of Sci ence 32 p Parviainen Arkkitehdit Ltd 2003 VTT BK5 Peruskorjaus Revision B Workshop drawing Espoo Finland 1 p Pietil inen A amp Merimaa M 2000 Mittaustekniikan perusteiden laboratorioty t 7 ed Espoo Finland Helsinki University of Technology Metrology Research Institute 108 p ISBN 951 22 5157 4 P
41. and at the site 3 for 12 hours The ground truth data was extracted from video recordings captured simultaneously with the counter data All counting data was sorted into 15 minutes intervals Yang et al 2011 Compared to the ground truth data the counting errors of the four counting periods were 20 5 22 1 24 8 and 14 3 with pedestrian volumes of 3 103 8 570 8 294 and 2 O11 respectively Varying this much no common correction factor for the count ing error could be used Additionally the Wilcoxon paired signed rank test was applied 21 to the sensor counts and the reference data suggested that there was a statistically sig nificant difference between them Yang et al 2011 Generally the IR beam sensors undercount pedestrians walking side by side or in groups Passing patterns of pedestrians at every test location varied randomly over time hence preventing the creation of unigue correction factors For enhancing the larger interval data a bivariate bootstrap sampling method using the 15 minute interval raw data was introduced as a sensor calibration procedure In the procedure four pairs of sensor readings X and ground truth readings Y were randomly sampled Based on the totaled values of sensor readings C and the ground truth values M an hourly counting error rate was calculated according to Equation 3 presented in section 2 1 The previ ous procedure was repeated sufficiently many times about 5 000 times to obtain a
42. and legs Cessford et al 2002 IR beam sensors have low power consumption and can hence be powered with batteries The sensors have a transmitter that emits a constant IR beam and a re ceiver that intercepts the beam When the beam is interrupted by an object a count can be registered with a data logger The transmitter and receiver can be in separate hous ings located on different sides of the monitored pathway or in the same casing In the latter case a separate reflector might be needed The path between a transmitter and receiver or a reflector has to be carefully aligned Hence the sensor can be prone to me chanical movements Sensors with two beams can be used to provide the walking direc tion information but the light beam sensors cannot count several pedestrians passing side by side Bu et al 2007 Assembling several dual beam counters one behind the other above the path perpendic ular to the walking direction it is possible to register several pedestrians walking side by side The main challenge with this setup is how to estimate the real number of pass ing people as only one person might trigger multiple sensors IR beam rows are suitable for wide corridors in large infrastructures like train stations Bauer et al 2009 Active IR sensor arrays illuminate the detection area with low power IR energy They detect the intensity of the light reflected from the target and are capable of classifying pedestrians and cyclists They still
43. ation is considered to be a major problem area in the retail industry Shopper guestionnaires are considered very unreliable as they are unable to reveal the details of the shopper movements spatial pattern and rely on the respondent s capability to remember their shopping experiences Pedestrian counting can be used to examine the shoppers circulation patterns to reveal the dead spots of the mall Plain counting however cannot be used to record continuous movements of individual cus tomers or customer groups Obtrusive tracking of individual shoppers tends to have an 41 influence on the monitored shopper s behavior whereas the unobtrusive observation of an unaware customer lacks the information of the basis of the customer s activities and can even be considered as unethical Brown 1991 The possibility to recognize and use geographical principles like social structure direc tion and the destination of pedestrian flows to determine the inter store relationships would be useful in predicting new retail sites As point form counting of pedestrian flows lacks any information about the customer trip patterns it is possible to use people flow modeling as a tool to translate the point counts into a map of networks For model ing a people flow counter has to be located at least at every pedestrian link of the ana lyzed network that is on a route without any junctions or turning possibilities except at its two ends As the number of
44. aturally this test data excludes the data of Ax is network camera sensor as it was not connected to the loggers If both counting directions were accounted as separate readings the datasets with 8 640 and 25 920 hourly readings in the six minute resolution were gathered for the direction insensitive and the direction sensitive sensors respectively For eight test sensors this makes at total of 155 520 six minute sensor readings A B Logger S N Sample start Sample stop e VI3004604T63190002 1 Takex kahvilaan 4 4 2011 8 54 4 4 2011 9 00 5 OPEN V13004604T63190002 2 Takex kahvilasta 4 4 2011 8 54 4 4 2011 9 00 1 OPEN VI3004604T63190002 3Cedec kahvilaan 4 4 2011 8 54 4 4 2011 9 00 5 OPEN VI13004604T63190002 4 Cedec kahvilasta 4 4 2011 8 54 4 4 2011 9 00 1 OPEN VI13004604T63190002 1 Takex kahvilaan 4 4 2011 9 00 4 4 2011 9 06 3 OPEN V13004604T63190002 2 Takex kahvilasta 4 4 2011 9 00 4 4 2011 9 06 7 OPEN V1I3004604T63190002 3Cedec kahvilaan 4 4 2011 9 00 4 4 2011 9 06 4 OPEN VI13004604T63190002 4 Cedec kahvilasta 4 4 2011 9 00 4 4 2011 9 06 6 OPEN VI3004604T63190002 1 Takex kahvilaan 4 4 2011 9 06 4 4 2011 9 12 4 OPEN VI13004604T63190002 2 Takex kahvilasta 4 4 2011 9 06 4 4 2011 9 12 1 OPEN VI13004604T63190002 3 Cedec kahvilaan 4 4 2011 9 06 4 4 2011 9 12 4 OPEN VI13004604T63190002 4 Cedec kahvilasta 4 4 2011 9 06 4 4 2011 9 12 1 OPEN VI13004604T63190002 1 Takex kahvilaan 4 4 2011 9 12 4 4 2011 9 18 1 OPEN VI13004604T
45. be adjusted separately to meet the corridor di mensions Any person appearing in the sensor s detection field causes a reflection of the IR beam at a distance different from the adjusted setting This operating principle makes the sensor more stable against light intensity variations compared to traditional beam 54 sensors Glossy or mirror like reflection backgrounds however can cause incorrect operation of the sensor The used custom Takex sensor employed an RJ 12 registered jack socket cabling for the supply voltage and the two output pulse channels Takenaka Electronic Industrial Co Ltd 2006 Position of detection 2 division photodiode object i Figure 32 Takex background suppression photo sensor left and the operation princi ple of the angle based distance recognition right Takenaka Electronic Industrial Co Ltd 2006 The operation principle of the Cedes TPS 210 two beam triangulation IR sensor Figure 33 left 1s also based on the angle of the received reflected light The sensor s limita tions are similar to the Takex sensor described above The Cedes sensor has two sepa rately adjustable sensor beams in the same compact housing with built in signal pro cessing and is insensitive to variations in ambient lighting conditions Figure 33 right A standard 4 pin M8 male connector is available for the power supply and the two pulse output interfaces The manufacturer reminds that applications of mult
46. came sufficient The assembly space limited by air ducts and cabling was on both sides quite narrow as can be seen in Figure 42 Hence two custom made mounting brackets were hence used to lower the sensors by approximately 300 mm from the ceiling If the sensors had been installed directly onto the concrete ceiling their fields of view would have been severe ly limited The Irisys and Sick sensors were installed on the caf side and the Eurotech and Axis sensors on the corridor side of the test site doorway A side view of the camera sensor placement is presented in Figure 46 and a view from the ceiling side in Figure 47 The ceiling mounting brackets are presented in Figure 48 Corridor Caf Uv E 0 N 4 PA O O A 244 cm 252 cm 1 5 cm 14 cm 1 2 cm Figure 46 Side view of the placement of the camera sensors Sensors F Irisys G Sick H Eurotech and J Axis The orientation of the Eurotech and Sick sensors is as shown in the inset pictures The placement of sensor mats is also shown D Emfit E Schmersal 67 Wall 72 5 cm Sem 86 cm 18 cm U El O O O O N Ww rc r Corridor Caf el14 cm Figure 47 Placement of the camera sensors as seen from the ceiling Sensors F Irisys G Sick H Eurotech and J Axis Figure 48 Mounting brackets for the camera sensors The aluminum clamps on the bracket behind are for the Eurotech sensor 68 5 3 Data logging The counting data from every test sensor was c
47. ce can be expensive Bu et al 2007 Video camera sensors use an image processing technique to subtract the static back ground track the remaining objects to determine whether they are pedestrians and to count them Variations in lighting conditions pedestrian clothing occlusions and shad ows can be challenging for computer vision detection Hence the cameras are usually mounted above the monitored area An extra benefit is the possibility to acquire a video capture for manual control simultaneously with the automatic counting Bu et al 2007 The video camera sensors are suitable for infrastructures of large people flows although they tend to undercount pedestrians during very high flow rates Bauer et al 2009 The needed image processing usually requires a lot of computing power Motion based de tection can efficiently reduce the number of false positive detections but it is worse in detecting stationary persons than the shape based method Stereo cameras are a good choice for accurate range measurements Chan amp Bu 2005 Automatic video monitor ing can also raise privacy issues Cessford amp Muhar 2003 Thermal IR camera sensors are often classified under PIR sensors as they don t emit any IR light but absorb heat emitted by objects Still they differ fundamentally from PIR sensors as they involve active image processing and can be used to detect multiple pedestrians simultaneously IR camera sensors can lose counting accurac
48. chnology is one of the many fields that widely uses automated detection and other sensor technologies Dharmaraju et al evaluated a selection of commercial sensors and counting systems potentially suitable for pedestrian and bicycle detection based on their ability to detect count and classify pedestrians and cyclists Other considered factors were the ease of installation mobility and performance under challenging weather conditions The eval uation was based on specifications provided by the sensor manufacturers Products in cluded in the survey were the Autosense II active IR sensor array the Traffic Vision video camera sensor the SmartWalk microwave sensor the Pedestrian Tactiles piezoe lectric sensor the IR 200 Dynamic Detector passive IR sensor and some unnamed pneumatic tubes and inductive loops Dharmaraju et al 2001 Cessford et al 2002 mention the knowledge of visitor numbers to be important for the management of conservation areas like national parks This information can be exploit ed in visitor flow and impact modeling and when planning visitor facilities services maintenance tasks and staff reguirements Visitor monitoring is described as a mix of different counting methods like direct observations on site counters and inferred 29 counts including visit registers and counts linked to fees and indicative counts like parking lot use As all visitor counting methods have their pros and cons the selection of a pr
49. cial pedestrian counting sensors promised to be high in accuracy are that also in price However assessing the suitability of different pedestrian counters for dif ferent locations without distinctive tests is difficult Yang et al 2010 tested the field accuracy of the EcoCounter passive IR counter and an IR camera counter provided by Traf Sys The EcoCounter Figure 19 left involved a dual direction counting technol ogy and an algorithm to avoid false counts caused by e g rain or vegetation move ment The EcoCounter s battery life was up to ten years and one year s data obtained in 15 minute intervals could be stored The sensor was installed on the side of the moni tored path Figure 19 right The Traf Sys IR camera counter Figure 20 left generated two adjustable imaginary counting lines that were used to determine the walking direc tion The counter needed a power supply and could store counting data in five minute intervals in addition to real time counting through a wireless transmitter The Traf Sys camera was intended for indoor use and had to be assembled above the pedestrian path Figure 20 right The optimal installation height of the sensor was about 3 5 m Yang et al 2010 Dual lens Wall Pedestrian Path EcoCounter Data Logger v m NN Figure 19 EcoCounter passive IR counter left and its test installation by Yang et al right Modified from Yang et al 2010 36 Mireless Transmitte
50. cs of the sensor were evaluated by measuring the IR radiation of a black body object placed in a thermal chamber in temperatures ranging from 10 C to 80 C The optimum chopping frequency was found to be 10 Hz and the sensor had sufficient sensitivity even at 100 Hz The time constant of the sensor was found to be about 5 2 ms making it faster in response speed than the commercially available reference pyroelectric sensor The test also results showed that decreasing the electrode size can be used to increase the pyroelectric sensitivity and the space resolu tion of the sensor Hashimoto et al 1999 The pyroelectric sensing system by Hashimoto et al could be extended to 360 degree presence detection e g for people counting in an auditorium The sensitivity of the de veloped 16 element sensor was more than double compared to the previous one with eight elements and one detection cycle using the 20 Hz chopping frequency and moni toring the 120 degree sector took 3 s In further experimental tests the smallest detecta ble distance between two persons standing next to each other was found to be 10 cm and the system was capable of sensing temperature differences bigger than 0 4 C To improve the resolution of the thermal image a data set was divided into nine parts and averaged The sensor system was also capable of detecting more special sources of heat like a lit cigarette or heated electrical wires and thus has potential in security an
51. culated by subtracting the reading of a day with no occupancy from the reading of a typical occupied day The modeled occu pancy was set to zero when the CO level was below a tolerance concentration of 50 ppm The estimates of the suggested sensor network were compared to the estimator relying only on the video camera people flow measurements This was obtained by comparing the estimates to a ground truth occupancy level that were got manually by analyzing the captured video frames The video cameras seemed to have a small posi tive bias resulting in a huge error at the time 18 00 The average estimation error of the model using only people flow information was about 70 at the building level and about 30 at the zonal level The errors of the suggested multi type sensor network were about 11 at the building level and about 21 at the zonal level The best esti mates were achieved by a suitable combination of different terms modeling different sources of information The suggested estimator was mentioned to be ideal for real time applications although its practical suitability and the impact of the sensor placement were still to be evaluated Meyn et al 2009 7 12 Sensors 6 O People count N 11 o TE Bary G G Figure 28 Layout of the sensor network for people traffic estimations by Meyn et al People count means a digital video camera Meyn et al 2009 50 4 Sensors of the test setup 4 1 Technical details of the test sens
52. d in a room with one person sitting in a chair at positions of 0 5 m 1 m 2 m 3 m 4 m and 5 m from the point on the floor directly under the sensor head It was found that beginning from 2 5 m the sensor output gradually decreased The sensor output was independent of the front side or back view position of the human body Between the standing and sitting positions however there was an effect on the output especially within 2 m from the sensor This was due to the fact that when stand ing a person s head and shoulders are closer to the sensor The sensor output value also decreased as the room temperature increased The room occupant counting performance was tested with 68 experiments achieving an accuracy of 85 3 Additionally a test to recognize people sitting in front of eight desks in a room 6 m x 8 min size was realized The sensor was located in the middle of the room and the number of occupants was var ied by up to ten persons the total number of test subjects being 389 The detection accu racy was 73 for an allowed tolerance of 0 persons and 97 for a tolerance of 1 person Yoshiike et al 1999 Mysen et al 2005 compared three ventilation strategies constant air volume CAV CO sensor based DCV and an IR occupancy sensor based DCV in 157 randomly se lected school classrooms in Oslo Norway The constant ventilation air velocity de signed for 30 people was 7 s per person and 1 s m due to material pollutan
53. d in the comparison tests mentioned here are listed in Table 2 Table 2 Pedestrian counting sensor types and manual counting methods discussed in the sensor comparison tests Test cn N S 5 a S O O on S N S O O L n zE E S RJR Aai ko a Sensor type or manual E on z counting method S N a Light beam IR or visible light X X Passive infrared PIR X X Computer vision video camera Computer IR vision IR camera X Micro wave radar Ultrasound Pressure e g switch tube X Manual counting on site X Manual counting from video recordings X X X X 2 2 Total number of sensors compared RE SRF Consulting Group Inc 2003 prepared a bicycle and pedestrian detection sensor survey for the Federal Highway Administration of the United States Department of Transportation and the Office of Traffic Engineering of the Minnesota Department of Transportation The project s primary goal was to identify possible sensor applications and evaluate the accuracy of non intrusive pedestrian and bicycle traffic detection methods through field tests A selection of similar projects and commercial detection systems was also listed and described The commercial sensors selected for the tests 31 suitable for both pedestrian and bicycle detection were the ASIM DT 272 combined passive IR and ultrasonic sensor the Diamond TTC 4420 IR sensor the MS Sedco SmartWalk 1400 microwave sensor and the ISS TCC Auto
54. d lighting automatically This way an increase of user comfort and energy savings can be gained People flow sensors can also be used for security purposes Hashimoto et al 1998 Yoshiike et al 1999 Hurych 2007 People flow monitoring and visitor counting has been widely utilized in nature parks and conservation areas around the world Cessford et al 2002 Arnberger et al 2003 Sensor technologies have also been widely tested for pedestrian and bicycle traffic counting and safety related presence detection Dharmaraju et al 2001 SRF Consulting Group Inc 2003 Chan et al 2005 Bu et al 2007 Diogenes et al 2007 For presence detection purposes the detection reliability of people in a certain area is often of more importance than counting every individual in the possible group or crowd Previous people flow sensor comparisons have been mostly accomplished in the context of out door counting or detection Larger scale sensor performance comparisons in indoor en vironments are however missing The objectives of this thesis were to design realize and document a test setup for a comparison of nine different people flow sensors Further the complete test setup was to be exploited for the collection of sensor data over a longer period of time together with limited time manual control counting Analysis of the gathered test data was left for future research Additionally the thesis objectives included a literature review on previous res
55. d safety applications as well Hashimoto et al 2000 Greene Roesel et al 2008 performed a review of commercial pedestrian counting sen sors based on different technologies Based on cost feasibility and commercial availa bility a PIR sensor provided by EcoCounter was selected for performance tests The device featured two pyroelectric IR sensors and an algorithm to avoid false counting generated e g by moving vegetation or sunlight Figure 5 Bounding of the detection zone with fixed objects could be used to further limit false counts The sensor was wa terproof and inconspicuous and its battery life was to ten years The device could be easily mounted to objects of different size and shape Greene Roesel et al 2008 Figure 5 EcoCounter dual PIR sensor Greene Roesel et al 2008 13 The test sensor was installed in three different outdoor urban locations at Berkeley Cal ifornia USA in May 2007 The locations were known to differ in their volumes of pe destrian flows and they had a proper place to mount the sensor perpendicular to the pe destrian passage Site number 2 is show in Figure 6 The test periods were four hours 12 00 16 00 at each site One contracted person manually recorded the pedestrian volumes using a hand held clicker and wrote the reading down every 15 minutes In the manual counting the direction of travel was ignored although the sensor was capable of sensing the direction A second contracted person
56. d would be needed and was a possible topic for further studies Installation skills and calibration costs also have to be taken into account when selecting a sensor and at heavily used urban loca 33 tions the sensor installation s resistance to possible vandalism is also a prerequisite SRF Consulting Group Inc 2003 Pedestrian presence detectors can be used in many ways to improve traffic safety They enable the observation of people awaiting their turn at crosswalks providing longer walk times across the street when needed restricting vehicular traffic s right to freely turn right when pedestrians are crossing the street and any general warning of motor ists Pedestrian counting sensors can additionally be exploited in crash exposure count ing Turner et al 2007 performed an evaluation test for pedestrian detection equip ment at a street intersection in College Station Texas USA The site was selected based on the pedestrian activity access to high speed communications level of pedestrian safety the possibility to demonstrate vehicle infrastructure integration applications in the future and pedestrian pattern simplicity Turner et al 2007 Based on previous evaluation tests and available product features the following pres ence detection sensors were selected for the tests MS SEDCO SmartWalk 1400 and ASIM IR 201 curbside sensors and MS SEDCO SmartWalk 1800 and ASIM IR 207 crosswalk detection sensors Sensors from MS
57. destrians and further to prevent colli sions Hence the goal here was more about detecting objects than counting them The technologies handled in the review however can just as well be used in traffic flow monitoring or counting pedestrians at an intelligent pedestrian crossing Different sensor technologies can also be combined to compensate their limitations This in turn can bring up some reliability issues since smart processing of the sensor data is necessary Based on the conducted review five sensors were selected for tests to be conducted in later studies the IBEO laser scanner the EVT 300 and the MS SEDCO SmartWalk 1800 microwave sensors the Irisys IR camera sensor and the SENIX Ultra 100 ultra sound sensor Chan amp Bu 2005 Bu et al 2007 give some pedestrian counting sensor examples and their prices in United States dollars USD Jamar Technologies USD 790 IR beam Irisys USD 1 400 IR camera EcoCounter USD 2 600 double passive IR EcoCounter piezoe lectric LASE Peco USD 9 000 laser scanner and Video Turnstile USD 1 230 vid eo camera The review by Greene Roesel et al 2008 includes the same sensor brands and models as the review by Bu etal 2007 The review by Bauer et al 2009 deals with the main applications limitations and typ ical accuracies of the people counting methods If the counting task is simplified limit ing pedestrian movement to a single file comparatively cheap and simple s
58. e 41 The reflector required for the Clas Ohlson single beam sensor was installed on the opposite side of the doorframe A side view of the precise placement of the IR beam sensors is shown in Figure 44 and a view towards the caf in Figure 45 The reflector of the Clas Ohlson sensor was acci dentally located higher than the sensor itself However this didn t affect the sensor op eration and the alignment turned out to be simpler by turning just the sensor The Takex sensor was placed on the wall at the rear of the doorframe This was done because when placed side by side the double beam sensors interfered with each other making their proper operation impossible 65 banal 14 cm 19cm 2 5 cm Cafe a Corridor 198 cm 125 cm 122 5cm Doorframe J cm Figure 44 Side view of the placement of the IR beam sensors Sensors C Clas Ohlson D Takex and E Cedes Doorframe Wall Reflector 122 5 cm 125cm Threshold Figure 45 Placement of the IR beam sensors view towards the cafe The brown line represents a niche in the doorframe Sensors C Clas Ohlson D Takex and E Cedes 66 All four camera sensors were mounted on the ceiling Here the installation height was a crucial factor since it determined the field of view of the camera A camera sensor in stalled too low would be unable to see the whole area to be monitored After removing the ceiling panels on both sides of the doorway the ceiling height be
59. e foot separation was for the Jamar TrafX and mond 58 50 and 40 respectively If the walking person stopped the false de tection rate was for the Jamar TrafX and Diamond 43 33 and 7 respectively For someone running in or near the detection zone the missed detection rate for the Jamar TrafX and Diamond was 67 20 and 40 respectively All sensors worked satisfactorily in the detection range test although the Diamond as a beam sensor per formed the best Tested mounting heights had insignificant effect on the counting accu 35 racy At the first of the free traffic test sites 15 of the pedestrians were in groups of two or more persons and 47 at the second site At the first site the overall counting error rate for the Jamar TrafX and Diamond was 34 11 and 7 respectively the total number of pedestrians being 470 for the Jamar and TrafX and 327 for the Dia mond The Diamond sensor was installed at a different location due to inadequate mounting facilities At the second free traffic site the overall counting error rate for the Jamar TrafX and Diamond was 36 26 and 24 respectively the total number of pedestrians being 967 for the Jamar and TrafX and 970 for the Diamond All the counting sensors consistently undercounted the pedestrians mainly because of people walking in groups However adjusting the raw counts was possible using a site specific calibration factor Turner et al 2007 Many commer
60. earch on people flow sensors and their applications The test sensors were selected in co operation with Teknovisio Ltd a Finnish company providing automatic visitor counting and logging systems for the retail sector The thesis work was carried out as part of the Customer movement work package of the larger 4D Space project that will be introduced in the following 1 2 MIDE 4D Space project Multidisciplinary Institute of Digitalisation and Energy MIDE is a research program of the Aalto University MIDE carries out long term projects that combine different areas of expertise of several research groups in the university and its partners in co operation The aims of the research program include strengthening of competitiveness of Finnish industry and research as well as related education MIDE s funding arose with the help of the Technology for life campaign that was organized in 2008 to honor the 100 year history of the Helsinki University of Technology now a part of Aalto Uni versity The MIDE research program includes 11 projects covering a broad selection of different research areas of engineering Aalto University 201 1b The MIDE 4D Space project focuses on the research and development of future retail administrative tools for shopping mall merchants and correspondents as well as new advanced services for their customers Exploiting user feedback and dialog is at the cen ter of the projects service development processes The
61. eceiver S sender P path and L for the dis tance between the clusters Chen et al 2008 25 2 8 Sensor calibrations Bertozzi et al 2004 developed a tool for evaluating pedestrian detection algorithms for video cameras Figure 15 Using the tool s graphical user interface GUI an operator could annotate every pedestrian s position and size in all frames of an acquired video stream This was done by drawing bounding boxes around objects body and head Based on this a description file File H in Figure 15 was created and compared to a similar kind of file made by the evaluated detection algorithm File A in Figure 15 Based on the comparison statistics the algorithm parameters and thresholds could be adjusted A key advantage of the tool was its integration to a development environment of vision algorithms which enabled directly check the effect of tuning a parameter on the detection performance The manual annotation procedure needed to be done only once for every video sequence For each video frame values of correct detections false positives and false negatives were calculated making it possible to evaluate the algo rithm s performance individually at every frame Bertozzi et al 2004 A test set of images was created by manually annotating about 1 500 images of video sequences from different city scenarios e g a parking lot and city center including dif ferent illumination and weather conditions The set was used to
62. ected to be displayed in hourly daily weekly monthly or quarterly intervals The results of the selected time span are additionally represented as line or bar graphs together with the previous counting period of the same length The sensor data could be downloaded in CSV format e g to create spreadsheets for closer analysis viis lt Service eva E eportsi Settings Language English Y Visitor statistics Data From Listing Report Type jig Time Region Comparison J Visitors Transactions _ Conversion rate Counting Group s TKK Sensoritesti Counting Group Visitor Count lt 2011 gt lt Week 14 gt Help for this report O e dh Day of Week Clas Ohlson amp 9754 7724 8884 9764 1 2394 6251 4294 5 9044 15 7 0 0 R Emfit 1 0004 7764 9324 9834 1 1944 6314 4624 5 978 13 7 0 0 H Eurotech amp 0 0 0 0 0 0 0 0 00 0 0 H Eurotech2 amp 5274 4064 4764 5044 6204 1244 2094 2 866 199 0 0 MN Irisys 3 5011 3764 4374 4994 5971 1204 2131 27434 16 1 0 0 NNN M Schmersal 1 0424 8134 9784 1 098 13514 9794 7004 6 9614 10 9 00 MN 3 3844 307 4 3614 407 484 814 166 2 1904 220 0 0 Mmm 5364 4334 4724 5174 6594 1324 2344 2 983 17 3 0 0 MN 5114 4014 457 4934 6254 1204 2094 2 8164 19 1 0 0 MEMEH Total 3 5 4764 42844 50019 54774 67694 28124 2 6224 32 4414 157 00 Figure 60 Counting report page of the Visit Server listing all the t
63. ection of commercial sensors and counting systems or a literature study of the related research The sensor character istics are often picked from the vendor or manufacturer specifications Pedestrian count ing sensor types and manual counting methods discussed in the reviews mentioned here are listed in Table 1 Only sensors and methods needing no equipment for the detected people are listed 28 Table 1 Pedestrian counting sensor types and manual counting methods discussed in the sensor comparison reviews Only sensors and methods needing no equipment for the monitored people are listed Review S en z e a 2 8 2 Ke C N S 2 N gt O E D B a a O a s 2 F re ll 5 O O B S 2 is Sensor type or manual E KZ K S D z counting method 5 5 5 a 5 a Light beam IR or visible light X X X X X X Active IR array Passive infrared PIR X X Computer vision video camera X Computer IR vision IR camera X X X Ultrasound xX Microwave radar X X X Laser scanner X X X X Mechanical e g turnstile gate X Pressure e g switch pneumatic tube X X Seismic and vibration X Piezoelectric X X Capacitive Electric field Manual counting on site X X X Manual counting from recordings X X Total number of sensor types reviewed Eea E e An Dharmaraju et al 2001 state that many of the recently emerged automated detection technologies have evolved with military applications Nowadays ITS te
64. ed accu racies over the test durations at different test sites Additionally the Wilcoxon signed ranks test was used to determine whether the sensors really undercounted at every test site Yang et al 2010 The total number of pedestrians at the test sites varied between 21 and 8 294 Both counters systematically undercounted the pedestrians and the overall error rates were 27 3 1 1 for the EcoCounter and 18 3 0 7 for the Traf Sys sensors The EcoCounter achieved an overall counting error of 5 3 or better only at a trail having a low pedestrian flow 270 pedestrians and at one of the crosswalk intersections 1 273 pedestrians Otherwise the overall error at the site was 14 3 or worse Differences in the EcoCounter s crosswalk performances might have been affected by possible sensor occlusions caused by pedestrians waiting their turn to cross the street The accuracy of the Traf Sys IR camera sensor lost to the EcoCounter only at the crosswalk intersection mentioned above its overall error being 14 6 This exception is probably due to pos sible pedestrians lingering in the detection zone of the Traf Sys sensor In addition to this only at a trail having high pedestrian volume the counting error of the IR camera 37 sensor was worse than 10 Neither of the counters however was designed for use at intersections The smaller overall error associated with the larger mean absolute error reflects the cancel out effect This situat
65. ena oli tehd katsaus ihmisvirtasensoreiden ja niiden sovellusten aikaisempaan tutkimukseen Kahdeksan testisensorin ulostulona oli niiden suunnantunnistuskyvyst riippuen yhdest kahteen pulssikanavaa Datanker ys toteutettiin yleiseen GSM verkkoon liitetyill dataloggereilla GSM loggereiden perusresoluutio oli yksi tunti mutta kuuden minuutin mittausdataa oli saatavilla tarkempia analyyseja varten Poikkeuksen muodosti yksi kamerasensoreista joka oli yhdistetty PC tietokoneeseen l hiverkon kautta ja tarjosi mittausdataa 15 minuutin resoluutiolla Testipaikaksi sopiva kulkuv yl valittiin Aalto Design Factorylta Sensorit kiinnitettiin ovenpieliin seiniin kattoon tai lattiaan valmistajien suositukset huomioon ottaen Koek ytt jen perusteella sensorit toimivat asianmukaisesti mutta testij rjestelyss havaittiin kuitenkin puutteita mm kaikkia sensoreita ei voitu asentaa t sm lleen samalle laskentalinjalle ja kamerasensorien asennuskorkeus oli rajoitettu Valmista testij rjestely k ytt en ker ttiin sensoridataa noin 30 vuorokauden ajan Lis ksi tehtiin reaa liaikaista manuaalista kontrollimittausta viiten erillisen yhden tunnin ajanjaksona samanaikaisen videokaappauksen kanssa Aiheeseen liittyv jatkotutkimus tulee sis lt m n mm t ss ty ss ker tyn sensoritestidatan analysointia sek sopivien sensorien hy dynt mist taloteknisiss sovelluksissa Avainsanat Ihmisvirta ihmisvirtasensori k vij laskuri se
66. endently decided whether to count the detection or not Clusters had to be syn chronized at least with the neighboring ones at the millisecond level Chen et al 2008 ae 993 E p T A NP geile ghia Aa N W Nal h a b Mica2 Ultrasound G Ultrasound Mote Receiver Sender Figure 13 Node pairing a and three node b ultrasound sensor cluster approaches for moving object counting by Chen et al Chen et al 2008 24 At an experimental setup a set of three adjacent ultrasound sensor clusters was installed on the ceiling of a laboratory at a height of 2 5 m in 30 cm intervals Figure 14 left The interval was selected so that two clusters could at most simultaneously detect the same object People of different heights passed the sensor network using seven different paths and 80 times per path Three moving paths were selected to be under the sensor clusters and four between or next to them Figure 14 right It was possible to separate two persons walking successively if the distance between them was at least 20 cm Tall persons were found to reflect more pulse energy than short ones The average number of reflected waves detected by every cluster was registered In most cases only one or two clusters detected one passing object The average counting accuracy including all the seven paths was 94 5 and 87 9 with and without the direction segregation respec tively Chen et al 2008 In a further test people moved in g
67. ensors can be used in automatic counting Traditional turnstiles can be replaced e g by IR beam or switching mat sensors For accurate counting however both of these sensors reguire that the spacing between consecutive pedestrians is sufficient Additionally moving back and forth in the sensor area causes overcounting and for the classification of the walking direction a double beam or double pad sensor is reguired In their sensor speci fications vendors often list the counting accuracy to be around 95 or even more with out any information about the flow rates or other conditions under which this can be 30 achieved Under normal conditions conditions the accuracy of any pedestrian sensor is usually well below the theoretical best value The relative error however decreases with temporal aggregation but possible systematic errors have to be taken into account separately Previous comparison studies of pedestrian counting sensors indicate that there is no one optimal sensor for all possible circumstances Bauer et al 2009 2 10 Sensor comparison tests People flow sensor comparison tests contain practical tests of two or more sensors whose performance is usually evaluated against manually collected ground truth data The test sensors can all be based on different operational principles or the test can in clude similar kinds of sensors from different manufacturers Pedestrian counting sensor types and manual counting methods handle
68. entering and exiting pedestrians at every link in a junc tion is then known the people flow model must solve the junction turning problem that is to determine the number of pedestrians moving from link A to links B C and D in Figure 23 Thornton et al 1991 Directly measuring the number of people turning at every link would notably increase the analysis time as every pedestrian would have to be followed through the intersec tion Additionally temporal variation would spoil the generality of the model Turning flows can be approximated by certain pedestrian behavior hypotheses but the simplicity of assumptions increases the error of the estimations In the absence of pedestrian flow characteristics some mathematical model can be applied The WONKA pedestrian movement modeling program presented by Thornton et al 1991 allowed the collected photographic counting data to be transformed into a network flow model In a full scale test pedestrian data was collected from a 57 link network The model was based on an apportion hypothesis that determined the turning flows from one link based on the rela tive numbers of pedestrians leaving a link junction through another link The program iterated and reduced the errors of reassignment until a total predetermined error thresh old was reached The obtained matrix showed how many of the pedestrians in each link turned at the connected links when the junction was reached If information about past and f
69. equency of the IR light sources of adjacent devices is varied to avoid mutu al influence between the sensors However the counting results of the multiple sensors are not automatically combined but are available on every device s pulse outlets or se rial interface The sensor is suitable for operation without ambient lighting The lower shielding class of the sensor s external cooling fan should also be accounted for when considering possible outdoor applications see Table 4 Sick AG 2011 firm firm MUUUUV UT eer VE TEES o Diffuser lenses A B A B infrared light sources a a A MN N EILA a f j i 1 SN N 4 y j d Figure 35 Sick time of flight camera sensor left and the walking direction separation principle based on two areas of the camera s field of view right Sick AG 2011 The Eurotech PCN 1001 is a direction sensitive people flow sensor featuring a stereo scopic camera system and built in IR illuminators Figure 36 left The device is pri marily intended for passenger counting solutions integrated above bus and train door ways The Eurotech sensor is hence suitable for outdoor installations and IR illumina 57 tors make operation in dim lighting conditions possible The sensor uses a stereoscopic view to analyze the shape height and direction of passing objects Eurotech S p A 2011 The Eurotech sensor has to be installed above and in the middle of the monitored corri dor and
70. er of pedestrians passing simultaneously in parallel through the measurement area Figure 2 below In the performed simulation test groups of one and six people could be recognized with 99 and 71 accuracies respectively The recognition rate of other group sizes was between these values The sensor readings were found out to be always less than the real number of pedestrians This was due to the fact that the occlusion of pedestrians in parallel is never complete and there were many conditions for parallel going persons It was concluded that a RBF neural network could be used to improve the counting accu racy of IR beam sensors and suggested that by adding to the measured value 3 of this value the measurement error in this experiment could be reduced further Song et al 2008 Classification Total Number and E m of People Recongnition Network Training Figure 2 Grid layout principle above and block diagram below of the IR light beam sensor based customer counting system by Song et al Song et al 2008 e Data Data Acguisition Processing 2 3 Active infrared sensor arrays An active IR sensor array is usually used to detect and classify motor vehicle traffic but it can be used for pedestrian and cyclist detection as well Motorcycles sorting rules can be used for bicycles but for pedestrians a new detection algorithm is needed When a target crosses a sensor s beam a change in distance is detected The cros
71. essive ly two persons in opposite directions four persons in pairs in opposite directions and 15 people walking freely the group sizes were excluded The counting accuracies of the sets were 97 85 86 64 and 84 respectively The system s response was also found to decelerate and miss fast moving persons in the lower sensor sampling frequen cies Other things having a negative effect on the sensor performance were the decrease of the separation between pedestrians sensor noise occurring in very slow movements differences in the sensors geometrical configuration an unexpected pedestrian size or motion pattern pedestrians stopping under the counter and temperature gradients Mathews amp Poign 2009 Schneider et al 2009 conducted a study on extrapolating short term manual pedestrian counts using counting data automatically collected with the direction sensitive Eco Counter pyroelectric dual IR sensors Manual counting was performed at 50 intersec tions in Alameda County California USA Of the 528 intersections along the state maintained main roads 30 intersections were selected based on variables of population density median income and proximity of commercial properties that have been found to correlate with bicycle and pedestrian traffic The remaining 20 locations were chosen randomly from the 6 938 non state maintained main and minor road intersections De pending on the estimates of intersection pedestrian act
72. est sensors The counting results of a certain sensor can be viewed by clicking the corresponding name on left red oval TI viisiit Service Language Enolish N Data From Report Type Sensoritesti v 05 Regions PLAN E Dayotweex EA 2011 Ki Week 14 Rimon JJ Time Region beares vaa RMisitors Transactions Conversion rate N B Update Summary Visitor count Lo II TKK Sensoritesti B Counting Group Visitor Count Takex Takex Kahvilaan lt 2011 gt lt Week14 gt lt Monday gt Hours Help tormis repon O dH am i mt F GO wD Mn ae 14 15 6 17 Z aia oew prevy fsoniud J Takex Kahvilasta a 1 0 0 0 0 0 2 5 27 26 51 49 42 41 44 37 32 26 15 9 7 10 0 0 4244 247 00 MI J Takex kahvilaan a 0 0 0 0 0 1 3 7 47 50 76 37 56 62 69 42 28 27 12 8 7 4 0 0 5361 468 0 0 M x 1 0 0 oF 0 1 5 12 74 76 127 86 98 1034 113 79 60 531 27t 174 14 14 oF 0 9601 362 0 0 A Visitors Total V Previous period Y E Current period Graph type Copyright Teknovisio Oy All rights reserved Figure 61 Single sensor one day report showing separately the flow volumes of both counting directions and the total flow rate in one hour intervals A The resolution and period of the report can be changed B and the data downloaded in CSV format C The graph below displays a comparison between the selected and the previous counting period of the same length The highest resolution countin
73. exploited inputs from a variety of sensor measurements and utilized historical data of building conditions pedestrian patterns and user preferences It was suitable for making building occupancy estimations that could be further exploited in e g security monitoring automatic ventilation lighting control and to accelerate possible safety evacuations The monitored building was di vided into zones that could be based in physical rooms or in groups of rooms The algo rithm used to estimate occupancy was based on the extended Kalman filter Meyn et al 2009 A test setup was built in an office building to evaluate the performance of the sensor network Three different types of sensors were used ten digital video cameras 12 PIR detectors and 15 CO sensors Video cameras were installed at the floor entrances and 49 in the middle of the corridors To obtain the walking direction the PIR sensors were placed in pairs between the video cameras The CO2 sensors were distributed in differ ent rooms and the floor was divided into 11 zones Figure 28 Additionally a priori knowledge of the building occupancy levels people flow rates and patterns seasons and times of day were used as algorithm inputs Meyn et al 2009 Due to the coarse data of the video camera and the PIR sensors a statistical error cor rection was applied Additionally in order to eliminate the day to day variations of the CO sensor readings a correction factor was cal
74. for the power supply and sensor interface communications Figure 17 Due to safety and security issues sensors and other measurement equipment had to be removed for the night SRF Consulting Group Inc 2003 J Autoscope Solo MS Sedco SmartW alk Existing Light Pole Detection Zone i ee 7 Diamond TTC Reflector yt M yi a aes Diamond TTC pr N ASIM DT272 N E 3M Microloop 4 D _ J Loop Detector 7 L Existing Handhole jo jf pf F F N Bicycle Trail Existing Cabinet Pedestrian Trail Figure 16 Sensor installation schematics of the survey by SRF Consulting Group Inc SRF Consulting Group Inc 2003 32 Sensors OARA Existing Light Pole Existing Cabinet Power Cables Power Strip Pedestnan Trail Figure 17 Test equipment schematics of the survey by SRF Consulting Group Inc SRF Consulting Group Inc 2003 Data collection was spread over two days one day was used for sample data and the other for official data collection The sample data was used to separately calibrate every sensor before the real tests by comparing the results with manual counting The test data was collected using the sensor s relay outputs Inductive Loop Detector a real time interface Autoscope Solo Diamond MS Sedco or software ASIM 3M Microloop Due to low pedestrian and bicycle traffic at the test site the test personnel walked jogged and rode bicycles during the off
75. g data that the Cognimatics TrueView counting software embedded in the Axis network camera sensor could provide was in 15 minute intervals For the sake of uniformity however the test data from the Axis sensor was collected in one hour intervals After logging onto the sensor through an Ethernet connection the data could be downloaded in CSV format An example of the one hour resolution data file structure of the Axis sensor is shown in Figure 62 20110103000000 00408CAD3D7F Axis 00408CAD3D7F 1 0 20110103010000 00408CAD3D7F Axis 00408CAD3D7F 1 1 20110103020000 00408CAD3D7F Axis 00408CAD3D7F 0 1 20110103030000 00408CAD3D7F Axis 00408CAD3D7F 0 1 20110103040000 00408CAD3D7F Axis 00408CAD3D7F 0 0 20110103050000 00408CAD3D7F Axis 00408CAD3D7F 0 1 20110103060000 00408CAD3D7F Axis 00408CAD3D7F 1 1 20110103070000 00408CAD3D7F Axis 00408CAD3D7F 0 0 20110103080000 00408CAD3D7F Axis 00408CAD3D7F 0 1 20110103090000 00408CAD3D7F Axis 00408CAD3D7F 2 4 20110103100000 00408CAD3D7F Axis 00408CAD3D7F 12 9 20110103110000 00408CAD3D7F Axis 00408CAD3D7F 27 15 20110103120000 00408CAD3D7F Axis 00408CAD3D7F 30 21 20110103130000 00408CAD3D7F Axis 00408CAD3D7F 31 34 20110103140000 00408CAD3D7F Axis 00408CAD3D7F 15 18 20110103150000 00408CAD3D7F Axis 00408CAD3D7F 16 12 20110103160000 00408CAD3D7F Axis 00408CAD3D7F 11 11 20110103170000 00408CAD3D7F Axis 00408CAD3D7F 13 11 20110103180000 00408CAD3D7F Axis 00408CAD3D7F 20 21 201101031900
76. h a width of 150 cm would have better covered the doorway of the test site location but choosing the wider mat would have increased the depth of the mat to 100 cm This again would have increased the possibility of the same person stepping many times on the sensor hence causing over counting Also the width of the Emfit sensor mat was slightly small for the doorway Verification of the proper operation of the sensors was done along with the installation and configuration of the devices using the digital indicators above the doorway The correct counting was checked using the patterns of one person and when applicable two persons walking side by side and one behind the other However after considering the above mentioned sensor installation challenges it can be concluded that testing the sensors in smaller groups would possibly be reasonable Selecting a narrower test site for the IR beam sensors and the sensor mats and a site with a higher ceiling for the cam era sensors the factors limiting sensor performance in the setup realized here could pos sibly be avoided The behavior of some of the passing people was also seen to have a negative effect on the counter readings Some people stepped over the mat sensors and during the control counting period couple of passers by also intentionally walked purposely back and forth in the sensor area as they noticed the manual monitors Despite the fact that the digital indicators placed above the doorwa
77. he temperature and sometimes they need to be built in an exist ing structure Cessford et al 2002 A combination of pneumatic tubes and inductive loops is considered capable of detecting and classifying of both pedestrians and bicycles Dharmaraju et al 2001 Seismic and vibration sensors register vibrations originating from a buried sensor Their benefits and drawbacks are basically similar to the pressure sensors but seismic and vibration sensors need very careful calibration at each site Cessford et al 2002 Capacitive and electric field sensors detect disturbances in the electric field between charged electrodes They are non contacting but their operation range is too low for practical pedestrian detection applications Chan amp Bu 2005 The accuracy of the traditional manual counting depends on the vigilance of single ob servers and the complexity of the monitored scene Manual observations however need only a little planning and are the most effective way for temporally limited small size counting especially when immediate results are needed Bauer et al 2009 Manual counting can also include descriptive data Cessford et al 2002 Video recordings ena ble repeated and even slow viewing of the counting events and hence increased accura cy but require hardware installations and a lot of time for viewings Bauer et al 2009 Camera hardware also needs power and can be subject to vandalism Manual field and video counting Is in
78. hen the three transducers were located on separate sensor boards Chen et al 2008 After the pedestrians passed the sensor cluster it separated the time stamps and calcu lated diagnostic values based on them At this point the designated threshold time value was taken into account As a reflected wave could possibly reach first the receiver fur ther in the direction of object movement a bogus detection of direction could occur This is could be alleviated by controlling the temporal overlap between the receivers detection periods Analyzing more than two movement directions was possible by set ting more ultrasound receivers at even intervals around one transmitter Using three receivers for instance would allow the analysis of at least six directions However in this case the analysis method would become more complicated Chen et al 2008 With one three node sensor cluster an area of about 40 cm wide could be covered The MOCUS network contained multiple adjacent three node ultrasound sensor clusters to completely cover the monitored area like a wide corridor A coordinator node was as signed in every cluster and an intra cluster analysis was used for object and movement direction separation While the same passing object could be detected by multiple sen sor clusters a distributed algorithm was used for inter cluster cooperation Based on the algorithm and detection information from its neighboring clusters every sensor cluster indep
79. ian movement patterns At any scale the aim is to produce a map of open spaces connect them through a certain relationship and perform the desired graph analysis of the network These kinds of representations seem to have an analogy with humans cognizance of spaces At the building or the small urban scale one possible option is a visibility graph analy sis VGA that can be used to find the inter visibility connections of the studied area This means representing the isovist of every point of the room with proportional colors to form an inter visibility graph Figure 24 left Using an agent based analysis a num ber of software agents representing pedestrians are released into the environment Each agent uses the previously determined inter visibility graph to get information for its choice of next destination and an agent analysis map is created Figure 24 right Add ing point form gate counts to the map the number of subjects passing each location can be found and further compared to a real people flow counting performed at the envi ronment The UCL Depthmap software package is free for academic use Pinelo amp Turner 2010 University College London 2010 M UCL Depthmap 10 Graph1 als M UCL Depthmap 10 Graph2 DEAR Y File Edit Map Attributes Tools View Window Help x Y File Edit Map Attributes Tools View Window Help 58 Cae SH g ist Cae SH Fee i Visibility Graphs R mA En v s a Visibilty Graphs RMA QE 8 S VGA Ma
80. ibasaki R 2008 Real Time Monitoring of People Flows and Indoor Temperature Distribution for Advanced Air Conditioning Control In th International IEEE Conference on Intelligent Transportation Systems Beijing China 12 15 10 2008 Piscataway NJ USA The Institute of Electrical and Electronics Engineers Inc pp 664 668 ISBN 978 1 4244 2112 1 Kivilahti A 2011 4D Space project review In The Annual MIDE Project Preview Seminar Espoo Finland 15 6 2011 pp 1 12 Leephakpreeda T amp Thitipatanapong R amp Grittiyachot T amp Yungchareon V 2001 Occupancy Based Control of Indoor Air Ventilation A Theoretical and Experimental Study ScienceAsia Vol 27 4 pp 279 284 ISSN 1513 1874 87 Mathews E amp Poign A 2009 Evaluation of a Smart Pedestrian Counting System Based on Echo State Networks EURASIP Journal on Embedded Systems Vol 2009 pp 1 9 ISSN 1687 3955 printed ISSN 1687 3963 electronic Melville S amp Ruohonen J 2004 The development of a remote download system for visitor counting In Siev nen T amp Erkkonen J amp Jokim ki J amp Saarinen J amp Tuu lentie S amp Virtanen E eds Policies methods and tools for visitor management Proceedings of the second International Conference on Monitoring and Management of Visitor Flows in Recreational and Protected Areas Rovaniemi Finland 16 20 6 2004 Helsinki Finland Finnish Forest Research Institute Vanta
81. ical dimensions or both Furthermore the selected site had no doors that could have possibly interfered with test sensors operation If doors were involved the sensors could probably be installed so that opening door itself would not block the IR beams or cause bogus movement in vid eo camera sensors While opening the door however people s movement usually stops or at least slows down before passing through it Hence the situation may cause possi ble unpredictable functioning of the sensors Moreover any temperature difference be tween the two sides of the opened door could cause airflows through the doorway thus hampering the operation of the IR camera sensor 5 2 Sensor assembly All the test sensors were installed at the selected test site following the manufacturer and vendor instructions The conceptual picture of the test sensor placements seen from the corridor side is presented in Figure 39 The view from the caf side is shown in Fig ure 40 and a close up view of the wall mounted IR beam sensors in Figure 41 The ar rangement of the camera sensors on the ceiling is seen in Figure 42 60 8
82. il ure The logger s readout sampling rate can be selected to between 6 minutes and 24 hours the default being 60 minutes The saved data includes the sensor value pulse or no pulse a time stamp and the channel number The internal memory capacity of the logger is 19 days to 12 years of data depending on the sample rate The 60 minute de fault sample rate allows storage of 197 days of counting data The four data loggers and power supplies for all test eguipment were placed above the corridor side detachable ceiling panel Figure 51 Teknovisio Ltd 2011a Power interrupted Beam 1 interrupted Beam 2 W in ae Out direction direction i l I I l l i i i I Output A PNP i n I I N I I Output B PNP I I i i I I Tip s tin tout Figure 49 Timing diagram of the Cedes IR beam sensor illustrating the principle of creating counting pulses of diverging people flow at separate output channels Cedes AG 2009 69 Figure 50 Visit Log 4000 GSM data loggers Figure 51 Visit Log 4000 GSM data loggers L and hardware cabling placed above the detachable ceiling panel on the corridor side One logger is missing from the pic ture 70 One Visit Log logger has four pulse channel interfaces with RJ 12 sockets for 6P6C modular connectors As direction insensitive and sensitive sensors needed one and two channels respectively altogether 13 channels and four loggers were needed for the data co
83. ill include the analysis of the collected sensor test data and the utilization of suitable sensors in building applications Keywords People flow rate people flow sensor visitor counter sensor comparison demand controlled ventilation AALTO YLIOPISTO LISENSIAATTITYON SAHKOTEKNIIKAN KORKEAKOULU TIIVISTELM soreita on saatavilla useisiin eri mittausmenetelmiin perustuen mm infrapunavalokennosensoreita sek video ja infrapunakamerasensoreita Ihmis virtasensorisovelluksiin kuuluvat mm laskentadatan ker minen asiakasvirtamallinnuksia varten tarpeeseen perustuva ilmanvaihdon hallinta sek turvaso vellukset Ker tyn sensoridatan laatu ja edelleen sen k ytt kelpoisuus ovat riippuvaisia k ytetyn senso rin tarkkuudesta ja luotettavuudesta N in ollen haluttuun sovellukseen on t rke valita sopivin senso ri Viimeisint tekniikkaa edustavat suorituskykyisimm t laitteet soveltuvat vilkkaisiin mittauskohtei siin mutta ovat kuitenkin yleens kalliita ja edellytt v t usein vakaita ymp rist olosuhteita Yksinker taisemmat ja halvemmat sensorit eiv t selviydy monimutkaisista tilanteista mutta ovat yleens ottaen sietokykyisempi erilaisia ymp rist olosuhteita kohtaan T m n lisensiaattity n p m r n oli suunnitella ja toteuttaa testij rjestely yhdeks n eri tekniikoihin perustuvan ihmisvirtasensorin vertailua varten sek ker t testimittausdataa valmista j rjestely k ytt en Lis ksi tarkoituks
84. imate space occu pancy levels and schedules building type and size pollutant generation rates and build ing costs of the control system Further heating and cooling energy savings are depend ent on the temperature and required rate of ventilated outside air Thus the cost effectiveness of each DCV application has to be assessed separately Inadequate per formance and high price may additionally decrease the attractiveness of some DCV ap plications Fisk amp De Almeida 1997 Indoor environment quality is dependent on the concentrations of dozens of volatile internally or externally generated gaseous and particulate pollutants in the air Most of them appear in measurable amounts Carbon dioxide CO2 is one of the occupant gen erated bio effluents and its concentration in the outdoor and indoor air is usually about 350 ppm parts per million in a volume and 500 2 000 ppm respectively At these concentrations CO gt is considered to have no direct adverse health effects but is an easi ly measurable surrogate for other occupant generated pollutants like body odors CO can be maintained at some predefined value usually under 1 000 ppm by ventila tion as there is a direct relation between the steady state CO2 concentration of indoor air and ventilation rates per person However it is evident that the CO level never sta bilizes at some precise level and peak concentrations are useless for defining the venti lation rates Thu
85. in a 15 m radius The LASE PeCo vertical laser scanner can cover a passage of up to 26 m wide and classify pedestrians according to their height Bu et al 2007 Top mounted laser scanners use two vertical height profiles to detect people and their walking directions With an installation height of 15 m up to a 26 m wide count ing line is possible Laser scanners costing several thousands of euros costing are some of the most expensive sensors for pedestrian counting Hence they are an attractive op tion for counting needs of corridor widths starting from about 5 m In addition closely spaced pedestrians tend to be undercounted even with a laser scanner Bauer et al 2009 Laser scanners are also suitable for object tracking but conditions affecting the visibility like fog and snow limit their operation range Chan amp Bu 2005 Mechanical counters include hinged boardwalks turnstiles gates doors and stiles They are inexpensive simple to build and maintain and can be installed to existing structures However their moving parts are subject to environmental burdens abrasion and vandalism Pressure sensors include pneumatic tubes sensor cables pressure pads and strain gauges They can be used to count both pedestrians and vehicles and can be concealed against weathering The pressure sensors may have adjustable sensitivity and logging delay to exclude false counts and they are low in power use However their operation can vary with t
86. inelo J amp Turner A 2010 Introduction to UCL Depthmap 10 London United King dom University College London VR Centre for the Built Environment 60 p Referred 21 11 2011 Available http www vr ucl ac uk depthmap tutorials introduction depth map v10 website pdf 88 Ruph Haller R amp Wernli M amp Filli F 2006 Visitor Counting with Acoustic Slab Sensors in the Swiss National Park In Siegrist D amp Clivaz C amp Hunziker M amp Iten S eds Exploring the Nature of Management Proceedings of the Third International Conference on Monitoring and Management of Visitor Flows in Recreational and Pro tected Areas Rapperswil Switzerland 13 17 9 2006 Rapperswil Switzerland Univer sity of Applied Sciences Rapperswil Institute for Landscape and Open Space Research Centre for Leisure Tourism and Landscape pp 72 77 ISBN 3 033 00935 2 Schneider R J amp Arnold L S amp Ragland D R 2009 Methodology for Counting Pe destrians at Intersections Transportation Research Record Journal of the Transporta tion Research Board Vol 2009 2140 pp 1 12 ISSN 0361 1981 Sick AG 2009 TVS100 People Counter Operating Instructions Waldkirch Germany 70 p Referred 18 5 2011 Available https www mysick com saggara pdf aspx id im 0033855 amp lang en amp page 1 Sick AG 2011 Laser Measurement Technology TVS100 Indoor Mid Range Datasheet Waldkirch Germany 4 p Referred 14 10 2011 Available ht
87. ing this a two dimensional thermal distribution could be obtained bringing the benefits of large and small sensing areas simultaneous ly using large areas improve the detection accuracy and small areas increase the sensor output for a passing person The chopping frequency affected the responsivity of the sensor elements and could be adjusted The sensor s accuracy can be affected by the presence of structural thermal sources like radiators The reliability of the developed sensor however could be improved by using simultaneously another sensor insensitive to thermal effects e g an ultrasonic sensor Hashimoto et al 1997 Figure 3 PIR sensor module with a motorized chopper with left and without a cover right by Hashimoto et al Hashimoto et al 2000 In the first phase one sensor was installed above a doorway its array parallel to the pe destrian movement direction Figure 4 The system also included a fuzzy clustering algorithm for occupant number and walking direction identification from the sensor s thermal image As the amplitude of the sensor elements signal was proportional to the object s IR radiation energy electrode size and sensor lens characteristics optimization was important for the detection accuracy The passing people were detected from the difference between the measured output caused by the object and the mean output of the plain floor The movement direction could be distinguished by detecting the slope of
88. ion occurred with both counters especially at the low volume trail site 270 pedestrians Yang et al 2010 The results of the Wilcoxon test showed that there was a significant difference between the sensor and the ground truth readings at the test sites of high volume A weak linear relationship was found between the pedestrian flow rate and the counting errors of the both counters The Traf Sys sensor overcounted slightly more than the EcoCounter The EcoCounter tended to overcount when people stayed near the sensor or walked slowly and to undercount when they walked side by side The reasons for over or undercount ing of the Traf Sys remained unclear Irregular counting errors however suggested that neither of the counters could be adjusted using the same correction factor for all count ing sites Yang et al 2010 Security and border checkpoints are the passenger flow bottlenecks at airports and de fine the capacity of the whole airport complex At checkpoints the service times and the number of available checkpoints determine the passenger throughput Using more checkpoints increases the staff costs but on the other hand improves customer comfort and potentially frees their time e g for shopping at airport stores Further knowledge about the checkpoint service time lengths can be a basis for staff scheduling Monitor ing people flow rates and possibly exploiting people flow modeling also improves the overall crowd control as pedestria
89. ions calculated the mathematical difference at each of them and determined if the difference was low enough for a block match The algorithm could be used to track several customers simultaneously Camera settings like focusing and tilt of the camera and lighting factors were critical issues af fecting the quality of the video capture Data from seven days was analyzed and differ ent algorithms for image segmentation and threshold selection reflecting the variable environmental conditions were used Manual control counting was made from the vid eo capture Newman et al 2002 From the results the customer patterns were seen to vary with the time of day and the day of week Compared to the manual control the video analysis program systematical ly undercounted the total daily customer readings at every observation point At the store entrance the counting error was 9 11 3 56 at the exit 5 16 2 12 at the checkout area 21 69 6 84 and at the customer service area 13 42 4 05 The cus tomer service area however also included one day of 11 11 counting error The reason for the greater number of customers arriving at the store than the number of cus tomers exiting could be due to the fact that some people only entered and left the store immediately or just visited the customer service desk Practitioners would like to know why there are hot spots with high customer level and cold spots with low customer level in a st
90. iple Cedes IR sen sors require a certain minimum gap between the devices in order to avoid interference between the sensors Although this seldom happens in people flow monitoring the sep aration has to be taken into account in the sensor test setups containing multiple IR sen sors at same location Cedes AG 2009a Beam 1 trigger level adjustment Integrated status LED Beam 2 trigger level adjustment 4 holes to mount the windows on housing 4 holes to mount the housing Figure 33 Cedes two beam triangulation IR sensor left and its front panel right Cedes AG 2009a 2009D 55 4 4 Infrared camera and video camera Sensors The Irisys IRC3020 thermal array based people counter is an IR camera sensor contain ing imaging hardware and signal processing in the same housing Figure 34 left The passive sensor uses a thermal array to passively absorb heat emitted by passing objects It is thus independent of ambient lighting conditions and its operation is possible even in total darkness The Irisys sensor is direction sensitive and is assembled overhead Thus it is capable of simultaneously counting multiple passers by entering the moni tored area The sensor has two relay outputs providing a counting pulse depending on the walking direction Wider corridors can be covered by connecting up to eight sensors together The used model of the Irisys IR camera sensor had a field of view of 60 de grees and was intended for indoo
91. ivity one to four manual observ ers were used Monitoring took place on one Tuesday Wednesday or Thursday as well as on one Saturday at each site Observations lasted two hours at a time between 9 00 11 00 12 00 14 00 or 15 00 17 00 Pedestrians crossing the road across any leg of an intersection were registered in 15 minute intervals and summed for the total intersection count At T type intersections pedestrians walking along the main leg of the intersec tion were also registered to make direct comparisons with four leg intersections possi ble Schneider et al 2009 Four EcoCounter sensors were installed in turns at 12 of the 50 selected intersections to get variations in pedestrian volume in different areas and under different weather condi tions One sensor remained at one place but other sensor locations were changed on a monthly basis Sensors were mounted on street signs or parking meters about 76 102 cm from the ground along the sidewalk within 30 5 m of the monitored intersec tion The operating range of the counters was 4 6 m and data was collected in one hour intervals Schneider et al 2009 Counts for every hour were averaged to get a weekly pedestrian volume profile for each sensor location The weekly profiles of 11 locations were further averaged to create a composite weekly profile This was used as a basis to extrapolate the two hour manual counts to estimate weekly crossing volumes of each of the 50 intersection
92. lay the images acquired from the two cameras of the sensor and the view below shows the current tracking process and the disparity map 74 The Axis network camera s installation height was set to 253 cm Again the installation height means the distance between the ceiling and the floor The sensor s counting zone and counting line was adjusted as shown in the live view scene of the monitored area in Figure 58 The proper operation of the test sensors was verified by having one person alone and two persons in single file side by side and in opposite directions pass several times through the test area The digital sensor indicators Figure 54 were monitored simulta neously As the Axis video camera was not connected to an indicator its operation was checked using the Cognimatics True View people counter software e CSMS Statistics Settings Help TirueviewaReo pleieountern Live view 21 seconds ago 1 0 1 minute ago 0 2 2 minutes ago 0 1 ins Out 25 211 COMMUNICATION Figure 58 The Cognimatics TrueView people counter software embedded in the Axis network camera sensor showing a live view of the monitored area The counting zone of the sensor is visualized with blue lines and the counting line as the red area between them 75 6 Sensor test data collection 6 1 On the data collection The completed people flow sensor test setup was used to collect counting data round the clock for 36 days between 22 3 26 4
93. left and border checkpoint right The red lines are manual counting lines and the green rectangles represent the switch ing mat Bauer et al 2011 The reference video camera was located on the ceiling at both checkpoints From the recordings at the security checkpoint bidirectional movement was manually annotated over three virtual lines L1 L2 and L3 in Figure 22 left Every person was followed along all the lines to be able to construct the number of people in the area between lines LI and L3 Totally 2 856 persons were followed At the border checkpoint the record 39 ing was annotated over two virtual counting lines L1 and L2 in Figure 22 right Bau er et al 2011 Comparing the cumulative sensor counts at the security checkpoint to the manually an notated bidirectional counts of 3 745 the IR beam sensor was seen to be more accurate yielding a result of 3 735 0 3 The reading of the switching mat sensor was 4 415 passengers 18 This was obviously due to the fact that at border checkpoints pas sengers usually proceed in one direction only but at security check points some passen gers possibly move back and forth to repeat some security procedure thus stepping more than once on the mat At the border checkpoint only 10 persons of 904 crossed the counting site in the wrong direction and the reading of the IR beam sensor was 205 12 6 of 182 and that of the mat was 1 017 412 5 of 904 The smaller total num ber
94. llection The loggers are also capable of providing the connected sensors 12 V DC power supplied through the RJ sockets The power supply options of test sensors can be found in Table 3 in section 4 1 Teknovisio Ltd 201 1a The Visit Log logger uses a 900 1 800 MHz dual band GSM GPRS data transfer to dispatch the sensor data to the Visit Service server of Teknovisio Ltd In this setup the data was sent once a day during the night when the people flow in the building 1s as sumed to be insignificant The sensor data was saved in server s data base where it is available through a web portal by providing a user name and a password The block diagram of the complete test data collection arrangement is presented in Figure 52 and the operating principle of the Visit visitor counting data transfer system in Figure 53 The Visit Service system employed the Finnish standard time and the same time was used to synchronize the Axis sensor with rest of the system Digital indicators were connected between the sensors and the loggers for quick check ing of proper sensor operation and for cumulative on site counting monitoring The two counting directions when available were presented on separate indicators The Clas Ohlson IR beam sensor used its own indicator All indicator readings could be reset on site independently of the data collected by the GSM loggers The counting indicators were mounted above the doorway on the corridor side Figure 54
95. lookup table of vectors containing the synthetic sensor readings true values and corre sponding errors Based on this information a correction factor for a new sensor reading C could be computed by selecting a group of vectors including readings C C where is a small value i e 5 10 Counting an average of the vectors error rates amp ave the calibrated sensor output Ceai could then be calculated C 5 Ccal The same procedure could be used for smaller interval data as well Yang et al 2011 The calibration procedure was validated using counting data from two field tests at the same location site 2 durations 11 h and 12 h as a training data set The data set was sampled 5 000 times and was set to 5 Counting data from the two remaining locations sites 1 and 3 was used as sensor readings to be calibrated The results showed that due to the calibration the overall error rate of the readings at site 1 dropped to 1 5 from the original 20 5 The corresponding improvement for site 3 was 4 1 from the error of 14 3 Hence the presented procedure could be used to reduce overall errors of the raw sensor readings using limited ground truth data Further research however would be needed to investigate the calibration procedure performance at different pe destrian facilities and with different sensor types Yang et al 2011 2 9 Sensor reviews People flow sensor reviews contain an evaluation of a sel
96. model left out This makes verification of the sensor specifications complicated as some manufacturers may have several models under one brand or technology Furthermore the tests were mainly per formed outdoors and hence larger scale indoor sensor tests were missing Applications of people flow sensors include visitor counting for commercial purposes like customer flow analysis automatic indoor environment control and safety applica tions The visitor counting applications have been widely studied in association with nature resorts like national parks Also the retail industry recognizes the value of the knowledge of customer flows Automatic environment control studies have so far main ly concentrated in exploiting CO sensing in demand based ventilation Some compari sons including people flow sensor related occupancy sensors however have been completed People flow sensors can also be used with other types of sensors to increase the reliability of occupancy level approximations in public buildings In this thesis a test setup for the performance comparison of nine different people flow sensors was realized The best possible test site location capable of accommodating nine test sensors was selected and the sensors were assembled obeying the manufacturers recommendations as far as possible As eight of the sensors were connected to the same data logging system the test setup in this respect enables a reliable comparison of the sens
97. mproved by increas ing the captured frame rate Chen et al 2006 Figure 11 Basic a d and inverse directional touching e f cases of pedestrian pattern merge split situations by Chen et al Chen et al 2006 Amin et al 2008 presented a people counter system combining the low cost Irisys IRI1001 IR camera and a webcam type video camera Figure 12 The heat continuous ly radiated by objects was detected by an IR camera and made visible in the form of a thermogram Hence with this combination it was possible to overcome typical prob lems of the plain video camera sensor like insufficient ambient lighting or bad contrast between colors of background and clothes worn by the objects The pedestrian detection principle was based on an ANN and background training of visual images Utilizing an ANN fully automating the learning and object classification processes was possible The image processing of the video and IR camera data were done separately and then further compared to increase the detection accuracy Amin et al 2008 22 For the ANN training data capture and counting experiments both IR and video camer as were mounted above the monitored area making sure that both of them shared the same scene Three different control scenarios were captured each with six experiments with variable lighting conditions and object positioning and movement Each experi ment contained about 150 IR or visual samples In
98. n densities in different areas are known Bauer et al 2011 Bauer et al 2011 evaluated the performance of two different people flow sensors in counting passengers and inferring checkpoint service times at an airport The first of the sensors tested was an IR beam sensor M18 RO20 PN with the M18 RF48 reflector manufactured by NAiS Figure 21 left The sensor costed less than USD 100 and could cover a path up to 2 m wide The second device was a switching mat sensor 750 x 1 000 mm in size manufactured by Tapeswitch and costing about USD 1 500 Figure 21 right The mat contained two integrated conductive plates that were connected when stepped onto and was originally designed for security purposes Both sensors were relatively cheap but non direction sensitive and hence could be properly used only in situations where a unidirectional single file movement and sufficient spacing between pedestrians were possible The size of the sensor mat should be limited to allow only one passenger occupying the mat at time and its place should be selected in such a way that a passenger will stand on it through the service time The selected test sites were a security checkpoint and a border checkpoint at Vienna airport in Austria The ground truth data was collected from simultaneously captured video footage Both of the checkpoints typically served small passenger groups and had a walking path up to 1 m wide The service time was defined as the time between
99. n of presence sensors included measurements of the overall error rate ac cording to Equation 3 given in section 2 1 the missed detection rate and the false de tection rate These were taken into account due to the fact that missed and false detec tions can possibly cancel out each other The overall detection error rate of the ASIM and the SEDCO presence detection sensors respectively was 9 32 and 11 39 the missed detection rate was 7 22 and 10 31 and the false detection rate was 2 16 and 0 13 The curbsides contained 2 752 and the crosswalks 750 observations The factor that decreased the correct detection rates at both the curbsides and crosswalks was the fact that all pedestrians did not walk clearly through the detection zones and the zones could not clearly be established for each sensor Additionally at curbsides some pedestrians not intending to cross the street were detected as doing so some still stand ing could but intending to cross were missed Deciding whether some pedestrians at the edge of the zone were inside or outside the zone was also sometimes difficult Also cars turning or stopping in the detection zone during pedestrian crossings sometimes trig gered the sensors or caused a prolonged detection Turner et al 2007 For a single walking pedestrian none of the sensors had any errors All three devices however had difficulties detecting people walking close to each other The overall error rate in case of the on
100. ncy of the United Kingdom UK promoting the conservation of wildlife geology and wild places of England now part of Natural England Altogether 20 Visit Log units and 22 Visit Mats were assembled Mats were placed at a depth of 10 cm under the path soil and the loggers with 12 V batteries were buried in a plastic container near the path The only visible parts of the installations were small GSM antennas attached to a tree or post One test location was later equipped with a small solar panel which proved to be a good solution for the counting system power supply The loggers collected the data in six minute intervals and sent it to Teknovio Ltd s server in Pargas Finland The loggers themselves had a buffer memory for storage of up to 60 days of data A non UK SIM subscriber identity module cards were found to be practical at remote wildlife areas as they can link to almost any GSM network to get a transmission path Melville amp Ru ohonen 2004 Figure 9 Pressure sensitive Visit Mat sensor before surface restoration by Melville amp Ruohonen Melville amp Ruohonen 2004 18 Dampness was found to cause some overcounting in GSM loggers that needed to be re booted Even complete flooding of logger containers occurred Some hardware calibra tion and software updates were also needed before correctly recorded counting and last ing connections to the Teknovisio s server were achieved No counter accuracy tests were performed b
101. nd riding a bicycle target distance 30 40 and 50 feet group spacing side by side 1 2 3 4 and 5 feet and sensor mounting height 3 4 4 5 and 5 feet For the group spacing test two persons passed the sensors successively hold a string between them to maintain the desired spacing Every test was repeated 15 times in both directions At the two re maining sites free traffic was counted Both of these sites provided high volume pedes trian traffic and one also a large proportion of groups of two or more people Video cap ture was taken for ground truth data for four hours and the video camera s clock was synchronized with the sensors The tests were monitored from a distance and the sen sors were seen to be undetectable to the pedestrians The Jamar and TrafX sensors re quired vendor supplied software to upload the counting data with individual time stamps but the Diamond s counting was an hourly read from the sensor s interface The TrafX sensor had the possibility to adjust the delay between the readings The group spacing however was found to be the biggest challenge when calibrating the sensors Direct sunlight at some installations did not appear to affect the counting Prior to han dling the video captured data detection zones were identified on the viewing monitor based on engineering judgment and the manufacturers specifications Two independent counts were made from the recordings Turner et al 2007 The evaluatio
102. need a separate detection algorithm for pedestrians and vehicles and are sensitive to adverse weather conditions Dharmaraju et al 2001 PIR sensors detect the IR radiation of a passing object Their properties are similar to the active sensors but PIR sensors consume less power Their detection range is de pendent of the IR characteristics of the object and its backgrounds Additionally sudden lighting changes can trigger false counts Cessford et al 2002 The sensors are usually based on pyroelectric technology and might have an adjustable temperature threshold Double sensors are capable of direction sensing but passive IR sensors can have diffi culties separating closely walking pedestrians Bu et al 2007 Piezoelectric sensor mats generate an electrical signal as a result of the mechanical pres sure applied by a person stepping on it The sensors require only simple signal pro cessing but a physical contact with the detected object is necessary Piezoelectric sen sors are often used together with other sensors e g IR beam sensors to improve the detection accuracy Chan amp Bu 2005 The piezoelectric sensors can differentiate be tween pedestrians and bicycles based on the signal characteristics Dharmaraju et al 2001 Several sensors can be connected together to cover a wider area and a timer can be used to eliminate overcounting caused by persons making multiple steps on the mat The installation of the sensors under path surfa
103. nsion 60 30m 60 x 30m Platform Platform O n 60 Rest Room Store Ticket Gate Platform Platform Measurement Area o Laser Scanner Figure 26 Layout of laser scanners positions with the corresponding measurement area by Katabira et al Katabira et al 2008 48 The tests took place during a station rush hour The people tracking accuracy of the sys tem was assessed as the ratio of pedestrian trajectories followed completely from en trance to exit In two and three minute sample periods the tracking accuracy was 81 2 and 88 9 respectively The station s actual air conditioning system was not used but the nozzles directions were instead simulated using an overlapped image of calculated temperature distributions and pedestrian tracking results The effectiveness of the sys tem was evaluated by the number of people receiving the airflow and the average re ceiving the flow Two situations were considered one with conventional static nozzles and another with turning ones Compared to the conventional system the tested con trolled system increased the total number of people receiving the airflow over 5 s during an observation period of ten minutes by 44 and the average time to receive the flow by 20 Katabira et al 2008 Figure 27 An example visualization of detected people flows by Katabira et al Kata bira et al 2008 3 3 Safety applications The sensor network by Meyn et al 2009
104. nsorivertailu tarpeeseen perustuva ilmanvaihdon hallinta IV Acknowledgements This licentiate s thesis was completed at the Aalto University Department of Electronics as a part of the 4D Space project of the university s Multidisciplinary Institute of Digi talisation and Energy MIDE research program I would like to thank the MIDE pro gram for financially supporting the project and further my thesis Professor Raimo Sepponen supervised my thesis work reviewing and commenting the ready thesis for which I wish to express my gratitude Likewise Pd like to thank professor Petri Vuo rimaa for acting as a second examiner for the thesis Laboratory manager Matti Linna vuo who acted as an instructor for the thesis work Pd like to thank for his valuable improvement suggestions for the thesis contents Thanks go also to university teacher Luis Costa who proof red the complete manuscript Teknovisio Ltd provided the vast majority of the test sensors and the data collection equipment to my usage for which I m appreciative Special thanks go to Teknovisio s project manager Rauli Laiho with whom the test setup described in this thesis was planned and completed in co operation I also want to express my gratitude to all my fellow researchers at the MIDE 4D Space project Special thanks go to James Culley Riikka H nninen and Sonja L tti for their valuable assistance in performing the defi nitely tedious manual control counting The Aalt
105. nth s worth of counting data stored as in and out counts in 15 minutes periods The setup and configuration of the video camera and the counting software is done using the same software The counting zone can be modified by dragging it into the desired shape in a live view screen Figure 37 right The proper operation of the sensor can be verified following the real time counting mode showing live video from the counting area and visualizing the latest counts Cognimatics AB 2010a 2010b The Axis video camera sensor has to be installed above the monitored area facing the floor The camera should not be tilted or placed above any moving object like an esca lator The Axis sensor cannot operate in total darkness and minimum illuminance of 80 Ix is recommended Very strong ambient lighting and sharp shadows should also be avoided in the camera s field of view Cognimatics AB 2010a COONS statistics tiveview BEU Hen Muay EvwReo ple Gountey Settings o W General lal M Counting zane ved Figure 37 Axis network camera sensor left and a view of the Cognimatics people counting software right Axis Communications AB 2010a Cognimatics AB 20100 59 5 Setup for sensor comparison tests 5 1 Test site location While considering possible locations for a test site for a people flow sensor comparison the Aalto Design Factory at the Aalto University Otaniemi campus appeared to be the most suitable fo
106. ntified as a person Adriano et al 2005 Chen et al 2006 presented a bi directional people counting system with one color vid eo camera based on a background suppression method and analysis of the object s hue saturation and intensity HSI histogram In the proposed counting scheme each pedes trian was given a typical value of pixels in the captured image and statistical data was 21 used to approximate the number of people in the current frame As each people pattern was labeled with a HSI color vector for tracking a bounding box was presented to sepa rate people wearing clothes of the same color Overlapping bounding boxes in adjacent images were hence judged as the same person Chen et al 2006 For the system testing a video camera was placed above the monitored pathway at a height of 4 2 m and a capture rate of 30 frames per second was used Controlled groups of one to five people with various direction combinations speeds and merge split ac tions Figure 11 passed the camera Some pedestrians carrying luggage were also used The counting accuracy of the system was found to be 85 7 100 It was noticed that when there were more than two persons in a moving pattern certain situations like fast moving people were hard to handle People walking very close to each other or in op posite directions or groups merging and splitting or making abrupt movements de creased the detection accuracy The accuracy however could be i
107. o Design Factory staff receives my thanks for co operation especially the Factory s machine shop for preparing the sensor mounting eguipment needed in the test setup Thanks belong also to all my colleagues at the Department of Electronics who have in any possible way contributed to this research work Further I d like to thank my friends and relatives for every cheerful moment that direct ly or in directly has sustained me I m especially grateful to my mother Leena and my in laws to be Helka and Seppo for their continuous support during my graduate studies Finally P d like to express my deepest gratitude to my fianc e Silke whose love and support made the completion of this thesis work possible Otaniemi 23 3 2012 Jussi Kuutti Contents ACKNOWLEDGEMENTS uma havin vta vkaleta okaVvkavakava uva utavukaka akala Nuku ka uka akakykakakalakko IV CONTENTS sauduuasdus da N a akat astu kuolaa ery Peer ann ey een ee avata V SYMBOLS AND ABBREVIATIONS sssecsccsssssocecssooocesesooccecessocceeesoococesssssocessoso VII 1 INTRODUCTION sis saamassatsss ssa aaaaaaasost ana nan knaLke SSK Un KN N NA KON KU SK NN KK N NA KONAN aR 1 1 1 BACKGROUND AND THESIS OBJECTIVES sumat enis E EE Ran t visa suuta haa KOSKEA Meese 1 12 MIDE 4D SPACE PROIECT euin e E N N E odo Nusa aa Ta 2 1 3 PEOPLE FLOW RATE AND PEOPLE FLOW SENSOR sesssssssssssssssseerrreeeeersrssssssstereereeeeeessssesssseserereees 3 1 4 STRUCTURE OFTHE THESI
108. o similar sensor modules and possibly auxiliary elec tronics and counting software The selection of a people flow sensor is based on the required accuracy reliability and practicality The price spread of the sensors is wide and as a rule of thumb sensors with good performance and high capacity are usually expensive Mathews amp Poign 2009 One common indicator that representing the counting performance of a people flow sensor 1s the counting error rate g x 100 3 where C is the counting reading of the sensor and M the manual ground truth counting result from the corresponding time span Yang et al 2011 Usability limitations of certain sensor types like privacy issues with recording video cameras or the possibility of real time monitoring have to be taken into account in the sensor selection as well For proper operation sensors might have some reguirements for the mounting location like minimum and maximum installation heights Ambient con ditions may also limit the sensor usability e g enough lighting is usually needed for video camera sensors and IR camera counters should be installed at locations where they are protected from strong temperature gradients and air flows Active optical light beam sensors can be based on IR or visible light They are small light in weight cheap in price temperature insensitive and have usually adjustable de tection delay to prevent overcounting caused e g by swinging arms
109. of the entire shopping center Des Rosiers et al 2005 Hence the shopping center management can specify the rental rates based on customer flow generation potential of a certain tenant at a certain location Eppli amp Benjamin 1994 The central places of the mall often have the largest customer flows and by implication the most valuable store locations Teirikangas et al 2010 The shopping center administration however quite seldom have a clear picture of the customer flow volumes usually only the number of people traveling through the en trances is known but the movement of customers inside the building remains unclear Using proper people monitoring technologies at selected strategic hot spots enables data collection for people flow modeling and simulations and further the claim of the infor mation of customer movements This information reveals the elements directing the movement of large people flows and the possibilities to channel them in a desired man ner Facts about the movements of large people masses connected to the information about individual customer movements offer a valuable tool for shopping center man agement It enables placement of individual stores in a way most beneficial to both the customers and the entrepreneurs This can be e g placing related shops at a suitable distance from each other or certain mall tenants in the proximity of the mall entrances Teirikangas et al 2010 The control of shopper circul
110. ollected in a centralized manner by regis tering counting pulses on available interfaces The only exception was the Axis video camera sensor that didn t provide pulse outlets but was connected to a PC situated in a nearby office through an Ethernet connection instead The Axis sensor had an embed ded Cognimatics TrueView people counter software that can be accessed using a web browser The principle of creating pulses of diverging people flow in separate output channels is illustrated in Figure 49 The example chart is of the Cedes double beam IR sensor If an object passes through both light beams in the in or out direction a pulse is sent to the corresponding output channel The idea is same with the camera sensors although the conclusion of the walking direction and the selection of the corresponding output chan nel are made based on image processing by the sensor software A single IR beam sen sor and mat sensors that are incapable of direction separation give all pulses at only one channel In the idle mode some of the sensors had the output pulse channel voltages at a high potential thus creating a zero voltage pulse in the case of object detection see section 4 1 Table 4 The counting pulse data collection was realized using Visit Log 4000 GSM data loggers Figure 50 A Visit Log logger can be used both as a local sensor data storage and transmission unit and is equipped with a 12 V DC back up battery in case of power fa
111. olved to form a comprehensive research ensemble Aalto University 201 1a Kivilahti 2011 Online grocery retailing Shopping center change processes Customer involvement Development of mobile Real Estate ENG services and applications Marketing ECON Design ARTS Retail vati Re al i r r on at ON Prototypes Processes Web Services SCI People flows 4D Space Sensor technologies People flow measurement Customer movement 3D modeling Pedestrian flow modeling i 3D models Individual customers Geomatics ENG Individual customers Sensor technologies Customer movement measurement and modeling Applied Electronics ELEC Transportation Engineering ENG Individual conditions Use of space and energy consumption Heating Ventilating and Air Conditioning ENG Figure 1 Research areas and corresponding research groups of the 4D Space project Modified from Kivilahti 2011 1 3 People flow rate and people flow sensor People flow rate is a physical quantity representing the number of people passing through a certain passage per unit time Brocklehurst 2005 p 21 defines people flow rate g as follows g vpw 1 where v is the mean speed and p the mean density of passing people and w is the pas sage width The eguation above can be assumed to concern situations where a people mass proceeds through a long passage of constant width In this thesis however only a volume of passing people at a point locati
112. olyethylene terephthalate PET film The sensor mat s upper surface is addi tionally covered with another shielding aluminum and PET layers Figure 29 left An acrylic adhesive layer is also available on one or both sides of the element film if need ed The sensor mat contains rectangular electrically connected sensor elements of size 290 mm x 300 mm and is manufactured in rolls of 580 mm wide Figure 29 right Efmit Ltd 2003 A change in the thickness of the piezoelectric film generates an electrical charge and creates a voltage across the electrodes In regular applications like people flow count ing involving sufficiently high changes in force the sensor film can directly feed the measurement instrument and no amplification or impedance matching is needed The sensor film thus needs no power for operation but in the test setup discussed in here it was equipped with an external module by Teknovisio Ltd providing counting pulses based on the sensor activation The sensor mat can be used to cover comparatively wide doorways and corridors However it cannot distinguish passers by stepping on the mat simultaneously nor recognize the walking direction Emfit Ltd 2003 32 Teknovisio Ltd provides a mat counter using the Emfit sensor film with an adjustable recognition sensitivity Hence it can be used to group visitors into e g children and adults based on their weight One pedestrian usually takes only one step on the sensor ma
113. on is unhazardous to the human eyes or skin Figure 35 left Sull looking directly into the IR sources is not recommended Based on the 3D measurements performed by the image processing of the sensor the device detects pass 56 ing people and their direction The camera s field of view is divided into two zones and people are counted only when travelling through both of them Figure 35 right The walking direction is discriminated by the order in which the zones are crossed The sen sor is capable of recognizing several people crossing the field of view simultaneously even if they were proceed in separate directions The height and width limits for an in dividual pedestrian can be defined the counting area restricted and possible interfering objects faded Sick AG 2011 The Sick video camera sensor has to be installed horizontally above the monitored area It should not be installed above reflective materials and direct sunlight on the device should be avoided The sensor is configured using the supplied software through a serial connecting interface The software features scanning and visualization of the current counting values The width of the optimal counting area w of the sensor depends on the installation height h W h 2 X 2 7 where w and h are in meters In the case of a low ceiling and wide corridor width the optimal counting area can be increased by installing up to 16 sensors alongside each other The fr
114. on is examined and a simpler form for people flow rate can be used N t g 2 where N is the number of people that passed and f is the selected unit time A people flow sensor is a device capable of measuring people flow rate and often also the direction of passing people It is also commonly called a pedestrian or visitor coun ter Pedestrian counting is based on the sensor s ability to detect single persons crossing a virtual counting line The sensor is triggered by physical signals caused by the pedes trians like visual appearance heat emission reflections of the body surface or pressure against the path Bauer et al 2009 A sensor literally means a device that is used to convert measurable physical or chemi cal phenomena or guantities e g mass temperature or number of objects to electric guantities like resistance or voltage pulses Pietil inen amp Merimaa 2000 p 57 Thus the definition of a sensor does not include possible measuring electronics or signal or data processing The people flow measurement devices used in the test setup realized in this thesis vary significantly in their operation principle and some of them have built in measurement software All except one of the sensors transform recognized passers by to voltage pulses in one or two outputs and are thus referred to as people flow sensors 1 4 Structure of the thesis A review of previous research on different people flow sensors and pedestrian c
115. oper counting approach is mentioned to be always a compromise between the accuracy and the practical capacity needed The sensor review by Cessford et al pre sented no particular commercial examples Cessford et al 2002 The depictions of available visitor counting sensors by Cessford amp Muhar 2003 are similar to the review by Cessford et al 2002 The reliability of the collected visitor number and pattern data is of major importance and key locations for wildlife park visi tor counting are identified as strategic bottlenecks For counting integrity and sensor calibration purposes the counting locations should include both permanent and tempo rary sites Tolerance of environmental conditions is of great importance for sensors used outdoors as temperature variations water and snow could cause a device malfunction Robustness against physical vandalism is also a major issue Cessford amp Muhar also mention that some people may intentionally interfere with counting sensors thus causing false readings In addition to improving the counting sensor hardware and software it is also essential to improve the integration of sensor calibration methods counting site methodologies and the linkage of the counting data to the comprehensive park monitor ing and management objectives Cessford amp Muhar 2003 Chan amp Bu 2005 reviewed sensor technologies that could be used to assist bus drivers in monitoring the presence and movements of pe
116. ore With suitable camera positions and algorithms the tested setup could be further used to track individual customers through the entire store This tracking data could be utilized to analyze the influence of the store layout and format on the customer 19 experience The shop interiors could then be re designed to create the desired circula tion and flows in the store Newman et al 2002 Heikkil amp Silv n 2004 developed a video camera based real time system for outdoor cyclist and pedestrian monitoring The requirements for the system in addition to high accuracy traffic counting included among other things the classification of pedestrians and cyclists route tracking and one week stand alone operation time Due to the de manding needs the monitoring was divided into real time tracking and off line analysis on separate computers Each object detection event and time stamp was saved to a re movable memory card and during the night time when lighting conditions were inade quate for monitoring the system was put to a sleep mode to save the battery Heikkil amp Silv n 2004 Using regular video cameras for traffic flow counting many visual clues like color his tograms coherent connected regions or blobs and object contours can be employed for image processing in the utilized software The image processing based object track er is often implemented using Kalman filtering Recently condensation and mean shift
117. ors Nine commercial people flow sensors all from different manufacturers were selected for the test setup The test devices can be sorted into three main categories based on their operation principle pressure sensitive sensor mats IR light beam sensors and IR camera and video camera sensors Of the nine sensors used one of the IR beam sensors and both of the pressure sensitive mats were unable to detect the walking direction of a passer by The core physical properties of the test sensors are tabulated in Table 3 and electrical and their ambient properties in Table 4 All the test sensors except the Axis video camera provided counting pulse outlets utilized for data collection in the test set up Thus the maximum counting speed capacity of a sensor was determined by the length of the provided counting pulse and the following idle or inactive condition during which a new pulse was impossible All IR and video camera sensors came with an indi vidual setup and counting software Table 3 Physical properties of the test sensors Brand Model Type Direction Dimensions Mass Installation Max corridor sensitive mm g height width cm cm Emfit L series Piezoelectric mat No 1 205 x 580 x 0 4 327 120 custom Schmersal SMS 4 Switching device mat No 1 000 x 500 x 14 8 500 100 Clas Ohlson DES 700 IR photocell with No 70 x 30 x 60 60 600 reflector Takex DL S202 Background suppression Yes 84 x 63 x 47 158 200 cus
118. ors However the test location had some noteworthy defects that could have a nega tive impact on the sensor counting performance Thus they should be taken into account as possible error sources when analyzing the sensor test data As the overhead camera sensors all needed to be assembled in the middle of the monitored passage they had to be installed back to back The door frame however was too low for camera installa tions and the camera sensors were assembled on the both sides of the doorway This caused a situation where all the sensors were in slightly different places within an area more than 1 m long in the corridor direction This can possibly be a problem in a situa tion where a person walks completely through the detection area of the first sensor and then turns back avoiding the rest of the sensors The installation directly to the ceiling would have guaranteed the proper installation height for the video camera sensors The air ducts on both sides of the doorway however would have blocked or limited the field of view of the sensors Thus the two mounting brackets were used to drop the sen sors by about 30 cm As a result the installation height of all camera sensors was re 80 duced close to or little below the recommended lower limit For example the installa tion height of the Axis video camera sensor was 244 cm while the recommended and theoretical minima were 300 cm and 250 cm respectively Additionally an emergency exi
119. ounting systems is presented in the section 2 The same section includes an overview of previ ous people flow sensor calibration studies and comparisons Section 3 presents a selec tion of applications for people flow sensors in retail demand based ventilation control and safety area The people flow sensors included in the comparison test setup realized in this thesis as well their essential characteristics are listed in section 4 The test setup itself is por trayed in section 5 including choosing the test site location disposition of the sensors and the sensor data collection system Section 6 contains a description of the exploita tion of the complete test setup for the collection of the sensor test data The realization of the manual control counting and comparison of the collected sensor data against the manual control is also presented The thesis conclusions are given in the section 7 to gether with intended future work 2 Previous research on people flow sensors 2 1 Sensor types People flow sensors or counters can be categorized by the detection method used Commercially available sensors include e g infrared IR light beam cells passive infrared PIR detectors video cameras infrared cameras laser scanners ultrasonic detectors microwave radars piezoelectric mats and switching mats Most of these are capable of sensing the movement direction of the monitored object or can be modified for this purpose by combining tw
120. p VGA Map i Drawing Layers ati Figure 24 An example of the Depthmap inter visibility graph left and an agent analy sis map based on it right Warmer colors represent higher visibility and agent flow rates Pinelo amp Turner 2010 43 3 2 Demand controlled ventilation Demand controlled ventilation DCV is used to achieve acceptable indoor air quality IAQ and energy savings by controlling the outdoor air ventilation rate The control can be based on some measured parameter or their combinations e g the value of an indoor pollutant temperature humidity or occupancy level or the ventilation of a room can even be controlled manually like using the room s light switch Ventilation recommendations are usually given in outdoor airflow rates per person and during un occupied periods the system can be turned to minimum flow settings or to totally shut off and this way gain notable energy savings Emmerich amp Persily 1997 Using a DCV offers potential energy savings in over ventilated buildings and improves IAQ in under ventilated ones If the minimum ventilation rate per occupant is reduced from 10 s to 5 s the estimated energy savings including the energy for fans heating and cooling vary from negligible to about 50 The largest energy savings can be achieved in buildings having high and variable occupancy levels The savings in venti lation energy however depend on among other things the local cl
121. p 23 28 ISSN 0007 070X Turner S amp Middleton D amp Longmire R amp Brewer M amp Eurek R 2007 Testing and evaluation of pedestrian sensors Research document College Station TX USA Texas A amp M University System Texas Transportation Institute 44 p 89 University College London 2010 UCL Depthmap Spatial Network Analysis Software Web document Referred 21 11 2011 Available http www vr ucl ac uk depthmap US Department of Transportation 2011 ITS Technologies Devices Infrared Detec tor Web document Referred 13 2 2012 Available http www walkinginfo org ped smart infred htm Yang H amp Ozbay K amp Bartin B 2010 Investigating the performance of automatic counting sensors for pedestrian traffic data collection In 72th World Conference on Transport Research Lisbon Portugal 11 15 7 2010 Lyon France World Conference on Transport Research Society pp 1 11 Yang H amp Ozbay K amp Bartin B 2011 Enhancing The Data Ouality of Infrared based Automatic Pedestrian Sensors Using a Nonparametric Statistical Method In Transportation Research Board 90th Annual Meeting Washington D C USA 23 27 1 2011 Washington D C USA Transportation Research Board pp 1 12 Yoshiike N amp Morinaka K amp Hashimoto K amp Kawaguri M amp Tanaka S 1999 360 direction type human information sensor Sensors and Actuators A Physical Vol 77 3 pp 199 208 ISSN 0924 4
122. project s interests involve em bedded software and service prototypes combining both physical location and context information The emergence of photogrammetric and other sensing techniques is stud ied and the project also handles the monitoring of people movement in public spaces Aalto University 201 1a The 4D Space project involves five schools and seven departments of Aalto University e School of Arts Design and Architecture Aalto ARTS o Department of Design e School of Economics Aalto ECON o Department of Marketing e School of Electrical Engineering Aalto ELEC o Department of Electronics Applied Electronics Group e School of Engineering Aalto ENG o Department of Civil and Environmental Engineering Transportation Engineering Group o Department of Energy Technology Heating Ventilating and Air Conditioning Group o Department of Surveying and Planning Real Estate Group Geomatics Group e School of Science Aalto SCI o Department of Media Technology Web Services Group Additionally research collaboration is done in particular with Teknovisio Ltd VTT Technical Research Centre of Finland and the Finnish Geodetic Institute Aalto Uni versity 2011a The research areas of the project and the corresponding research groups are presented in Figure 1 Although the research topics are marked as a brief of a certain research group or groups the 4D Space project strives for a collaboration of all of the parties inv
123. r F Pedestrian Path Height MIU Data Controller Figure 20 Traf Sys IR camera counter left and its test installation by Yang et al right Modified from Yang et al 2010 The field tests of the sensors took place in seven locations in New Jersey USA The test sites were selected based on the volume of pedestrian flow sensor mounting facilities and location accessibility The sites included three trails one sidewalk two crosswalks and one pedestrian bridge The test periods varied between 6 and 14 hours Both the counters were installed according to the vendors specifications and the ground truth data was extracted from a video camera capture All collected data resolution was con verted to one of 15 minute intervals Yang et al 2010 The sensor counting results were analyzed by calculating for both sensor data a mean absolute error by averaging of the absolute values of the 15 minute interval error rates at a single test site counting the error rate 1s defined by Equation 3 given in section 2 1 With this indicator the cancel out effect of possible over and undercounting is cor rected and the differences in data collection durations at different sites are taken into account Also an overall error rate was calculated by taking the relative error of the sums of 15 minute counter readings and the sums of the corresponding ground truth readings at a single test site The overall error was used to compare the aggregat
124. r the purpose As a hands on working environment for many product development courses student projects and a venue for various day conferences and evening events it had a high number of visitors in addition to its permanent personnel Thus it seemed to be able to provide our test setup higher people flows than other possi ble university locations like corridors of regular research laboratories Moreover get ting permission to install a broad selection of test eguipment was more straightforward at the Design Factory compared to other busy public areas in the university campus Measurements of corridor width and ceiling height were made at several locations at the Aalto Design Factory using an ultrasound distance measurement device The floor plan of the Design Factory s main floor with selected measurements is presented in Figure 38 The second floor and the basement were not considered since most of the building s activity takes place on the main floor A location for the test setup was selected at an entrance between the lively Design Fac tory caf and a busy corridor where many of the building s meeting rooms are situated The passage had a comparatively high people flow rate in general and the corridor width and height were suitable enough for the sensor installations The selected test set up location is marked with a red circle in Figure 38 Other Design Factory locations examined were lacking high people flow rates suitable phys
125. r use free from rapid ambient temperature or humidity changes InfraRed Integrated Systems Ltd 2010 The configuration of the sensor is done using a separate setup module and computer software The module connects to a computer via a serial or USB universal serial bus port The sensing area is square shaped its width being approximately equal to the sen sor mounting height In the setup mode two counting lines are visible one for each counting direction The position and shape of the counting lines can be modified in a drag and drop manner Figure 34 right The operation of the sensor is additionally visualized with one or more circles followed by a solid line representing passing per sons and their recent paths respectively When a person crosses a count line in a certain direction a pulse is sent to the corresponding relay output The Irisys sensor can be con figured with various special counting modes for instance ignoring possible U turns InfraRed Integrated Systems Ltd 2010 Figure 34 Irisys thermal array based people counter left and an example of its setup view with red and green counting lines right The yellow circle is a detected passer by and the following black line their recent path InfraRed Integrated Systems Ltd 2010 The Sick TVS100 video camera sensor utilizes time of flight technology The detection area 1s illuminated by IR sources that use radiation having a wavelength of 850 nm that in normal operati
126. rating the tracked people positions and temperature distributions the degree of ventilation demand could be eval uated The obtained pedestrian trajectories could also be used to extract principal people flows for more ecological ventilation control Katabira et al 2008 For test purposes eight horizontal single row type Sick LMS200 laser scanners with an eye safe 905 nm laser beam were installed at a railway station in Japan The scanner type s maximum detection range was 30 m and distance measurement accuracy 4 cm Scanning planes of the devices were 180 degrees with angle measurement accuracy of 0 5 degrees The used scanning rate was 37 5 Hz and the devices were installed at the human ankle level The scanner positions at the station hall are shown in Figure 26 The scanners were connected to wireless MICA2 network nodes capable of organizing an automatic ad hoc network and relaying the data packets from node to node They also included an on board sensor for ambient temperature measurements with one second sample rate Altogether 12 wireless nodes were installed at heights of 160 180 cm The obtained range and temperature data was further sent to a central computer that ana lyzed the people flow information The computer calculated an optimum direction for the airflow and controlled the air conditioner eguipped with turning nozzles A visuali zation of the calculated people flows is presented in Figure 27 Katabira et al 2008 Dime
127. realize a test setup that can be used to compare nine people flow sensors based on different technologies and to collect test data by exploiting an existing setup Addi tionally the literature is reviewed for previous research on people flow sensors and their applications Eight of the selected test sensors provided either one or two pulse channel outputs depending on their direction sensitivity The data collection was implemented using data loggers connected to a commer cial server through a public GSM network The basic resolution of the GSM loggers was one hour but six minute data was available for closer analysis An exception was one video camera sensor that was connected to a local PC via an Ethernet connection and provided counting data with a resolution of 15 minutes A suitable doorway for a sensor test site was selected at the Aalto Design Factory Sensors were mounted on the doorposts wall ceiling or fixed to the floor according to the manufacturers recom mendations The operation of the sensors was verified and found out to work properly There were however certain drawbacks with the test setup all the sensors could not be installed on the same counting line and the limited mounting height for the camera sensors The completed test setup was used to collect sensor data for about 30 days together with five separate one hour manual on site con trol counting sessions including simultaneous live video capture Related future work w
128. resolution To get a sensor counter 26 calibration factor the number of manually observed visitors M was divided by the num ber of automatically registered visitor C A consistent calibration factor cf could be cal culated from the mean of the sum of every hour s value M ce mean 2 4 The corrected reading could thus be calculated by multiplying the sensor reading by the correction factor Additional tests were made to find the sensitive area of the slab sen sor how groups must pass the sensor to be counted correctly and the effect of the length of the step taken and the covering material on the counting Three test sites with different soil materials were used and group sizes and forms were varied Ruph Haller et al 2006 Except for overcounting in some cases of few visitors the sensor counters were found to systemically undercount the visitors Possible causes of undercounting could be visi tors stepping over the sensors passing very slowly over them turning back walking side by side or very close to each other Also insufficient time synchronization of clocks of automatic counters and manual observers and in winter even the frozen soil could be sources of the counting error The calculated two day overall sensor counting accuracy of the sensors in different locations was 44 8 62 7 The variation between sensors could be due to other circumstances like wrongly installed sensor slabs sensor burial depth path
129. reviations Symbols C Sensor counting result Ccal Calibrated sensor counting result C Indoor air carbon dioxide concentration G Supply air carbon dioxide concentration Cf Calibration factor G Carbon dioxide generation rate per person h Height Manual counting result N Number of people Os Supply air ventilation rate g People flow rate t Time v Mean speed of people w Width Wo Width of the optimal counting area Interval delimiter E Hourly counting error rate Os Average counting error rate p Mean people density VI 1 s l s PP PpMy min 1 s min l s m s To Vo 1 m Abbreviations Aalto ELEC Aalto ENG Aalto ECON Aalto SCI Aalto ARTS Al ANN APeC ASHRAE AVI CAV CCTV CO CSV C DC DCV ESN EUR GSM GUI HSI Hz IAO VII Aalto University School of Electrical Engineering Aalto University School of Engineering Aalto University School of Economics Aalto University School of Science Aalto University School of Art Design and Architecture Aluminum Artificial neural network Automated People Counting from Video American Society of Heating Refrigerating and Air Conditioning Engineers Audio video interleave Constant air volume Closed circuit television Carbon dioxide Comma separated values One of the popular general purpose programming languages Direct current Demand controlled ventilation Echo state network Euro currency Global sy
130. roups of one two side by side and two one behind the other The paths were set to be free and crossing several sensor fields Replicate counting occurred at very low velocities about 0 25 m s and when the path crossed several clusters Differentiation of a slowly moving single person and two closely mov ing successively persons was impossible due to the possibly similar detection sequenc es The accuracy for a single person detection was 90 and 50 70 with and without the direction segregation respectively Fast moving short persons were missed most easily As a result of the different sensitivities of the ultrasound receivers the used in ter cluster algorithm was found to overcount or undercount on some clusters This counting error had to be minimized by sensor calibration based on the average number of reflected waves from every cluster during the test of seven regular paths Adding the calibration factors to the results of the latter test the portion of missed and replicated counts reduced leading to an improvement of the counting results It was also men tioned that ultrasound sensor boards were power consuming and in practical applica tions it could be reasonable to schedule the counting to needed periods only Chen et al 2008 RO RO Red 4 SY S S i Mor Ae R T U Y L L Py PaP3P4P PsP Figure 14 Ultrasound sensor cluster placement left and pedestrian moving paths right by Chen et al 2008 R stands for r
131. rtation system ITS applications Different microwave sensor technologies can be integrated for more versatile operation They demand simpler signal processing than computer vision and can operate in variable environmental conditions Microwave sensors can also be hid den behind materials that are permeable to radio frequency signals and can be used to classify detected objects based on power spectrum of the reflected signal Chan amp Bu 2005 Microwave sensors are often used for pedestrian presence detection at intersec tions Dharmaraju et al 2001 Ultrasonic and microwave radar sensor solutions have not yet to the author s best knowledge shown satisfactory functioning in real world pedestrian counting applications Bauer et al 2009 Laser scanners emit IR laser pulses and detect reflections from the objects using the time of flight method One scanner can cover the entire 360 degree viewing angle Ap plying a procedure similar to image processing a high resolution image of the surround ings can be obtained Chan amp Bu 2005 Depending on the device the scanning can be performed horizontally or vertically The distance detection accuracy of the scanners are in the order of a centimeter and depending on the scanning frequency the azimuth an gel accuracy is between 0 25 1 degrees Pedestrians can be classified by the characteris tics of their moving legs The horizontal laser scanner manufactured by Sick AG can detect objects with
132. s The time of day and day of week were accounted to get an accurate estimate Also location and weather based adjustment factors were created While manual counting at each site took place in two time slots two weekly estimates could be averaged to create a final weekly pedestrian volume estimate Typical weekly volumes could be further multiplied by a monthly factor to develop annual estimates Preliminary annual volume estimates at each intersection were then compared to police reported pedestrian crashes Schneider et al 2009 The counting error of the EcoCounter sensor was found to be 19 9 in previous stud ies The result was consistent with this study during high volume gt 400 pedestrians per hour and low volume lt 100 pedestrians per hour periods Hence the sensor under 16 counting was unrelated to the pedestrian flow rate Many counting errors were noted due to bicycles and people standing in front of the counters but also for people walking back and forth in the monitored area These errors were removed from the data by com paring sensor counts to the manual ones However the manual counts could not be fur ther verified due to lack of simultaneous video recordings Additionally one counter installation was delayed due to permit processing and its data was missing from the analysis Schneider et al 2009 To be able to extrapolate short counting results to longer periods accurate adjustment factors would be needed
133. s the CO2 generation rate would be more reliable surrogate for room occupancy level instead of the instantaneous CO concentration Fisk amp De Almeida 1997 Installation costs of a DCV control depends on the current ventilation system If digital ventilation controls already exist CO2 based control requires only reasonable extra in vestments Fisk amp De Almeida 1997 Non uniformities in the building s air distribu tion and occupancy level can make comprehensive placement of CO sensors demand ing The sensors should be placed in rooms at mid level and away from doorways radi ators windows people and air inlets Another option is to locate them in the ventilation return ducts CO2 control is suitable for spaces like auditoriums cinemas and educa tional facilities where the occupancy levels are more variable and unpredictable where the occupancy peaks are fairly high the pollutant of non occupant sources is low and where heating or cooling is needed for most of the year Emmerich amp Persily 1997 44 Indoor air however also contains variable amounts of other non occupant generated pollutants than CO2 Thus monitoring of the CO level alone guarantees an acceptable IAQ and offers benefits when occupant pollutants dominate the building pollutants Emmerich amp Persily 1997 Non occupant generated pollutant sources include for in stance building materials furniture cleaning agents and the pollutants included in the
134. s A Physical Vol 79 1 pp 46 52 ISSN 0924 4247 Heikkil J amp Silven O 2004 A real time system for monitoring of cyclists and pedes trians Image and Vision Computing Vol 22 7 pp 563 570 ISSN 0262 8856 Hurych D 2007 Counting People in Videos In Martinovic J amp Dvorsky J editors Embedded Systems amp Microsoft Technologies Student Research Papers Ostrava Czech Republic VSB Technical University of Ostrava Faculty of Electrical Engi neering and Computer Science pp 15 24 ISBN 978 80 248 1650 0 InfraRed Integrated Systems Ltd 2009a IPU 40188 People Counter Mounting Height Graph Issue 1 Technical note Northampton United Kingdom 1 p InfraRed Integrated Systems Ltd 2010 Irisys IRC3020 Indoor People Counter With Relay Output Datasheet Northampton United Kingdom 2 p Referred 19 10 2011 Available http www irisys co uk people counting our products datasheets K A Schmersal GmbH 2010a Operating instructions Safety mats SMS 4 amp SMS 5 Rev C Wuppertal Germany 8 p Referred 18 5 2011 Available http www schmersal net Bilddata Si takti Pdf sms4 5 bedien en mrl sms4 5 en pdf K A Schmersal GmbH 2010b SMS 4 SMS 5 safety mat Product information Rev 1 Datasheet Wuppertal Germany 5 p Referred 14 10 2011 Available http www schmersalusa com cms17 opencms media loader id 49 amp ty pe pdf amp download true Katabira K amp Zhao H amp Nakagawa Y amp Sh
135. s connected to wireless acquisition units and a base station for data processing The sensors were in stalled overhead to avoid errors encountered in side mounting and in two rows to enable the direction sensing Figure 7 One counter covered a passage of 120 cm and several counters could be connected together The sensor counts could be predicted from data patterns using a machine learning technique An artificial neural network ANN name ly the echo state network ESN was used as the machine learning algorithm due to its efficiency with noisy and error containing training data For the network training a da tabase with various motion patterns was created Mathews amp Poign 2009 2C interconnection Tmote sky wireless sensor node PIR sensors TTA N I2C I O Sa AMD ee TtT expander Base station s processor F 4 a a Traffic flow direction Listener node Figure 7 Direction sensitive pedestrian counter system based on a PIR sensor grid by Mathews amp Poign Mathews amp Poigne 2009 The system was tested for the ESN and pedestrian counter performance Using 565 mo tion patterns for training and 465 for testing the network had a detection accuracy of 99 with no noise The performance deteriorated when the noise level was increased The performance of two prototype counters installed in series was tested using different controlled sets of people moving in both directions one person two persons succ
136. scope Solo video camera sensor Additionally the 3M Microloop and the Inductive Loop Detector magnetic loop sensors only for bicycle detection were also used SRF Consulting Group Inc 2003 The test site was a pedestrian and bicycle commuter facility with physically separated pedestrian and two lane bicycle trails The Inductive Loop Detector sensor was already installed under the bicycle lanes and its count station with a cabinet were used to pro vide power and a shield for the data collection equipment of the rest of the sensors The 3M Microloop magnetic loop sensor probes were installed under the bicycle trail pave ment and the remaining sensors were mounted on a light pole in the strip between the bicycle and pedestrian trails Figure 16 Installation heights for the ASIM Diamond MS Sedco and Autoscope Solo sensors were 3 m 4m 10 m and 12 m respectively The pole mounted sensors were tilted to face the same detection zone The reflector needed by the Diamond sensor was attached to a wooden stake on the opposite side of the trail Other necessary equipment were PC personal computer laptop computers for sensor configuration and real time monitoring an automatic data logger Peek ADR 3000 for counting data collection from the relay equipped Inductive Loop Detector sensor a television for sensor configuration and real time monitoring a video cassette recorder for establishing a permanent record of test activities and a terminal panel
137. sing of con secutive beams is used to create the object s view profile and an algorithm classifies the target Dharmaraju et al 2001 tested an overhead installed IR array capable of measur ing the size of objects at an outdoor trail in Amherst Massachusetts The United States of America USA The collected data included the total number of pedestrians and cy clists using the trail Manual counting was performed simultaneously and continued until enough data was gathered to achieve 95 confidence level with an error margin less than 2 about 601 observations It was noted that some single pedestrians or pe destrian groups were classified under one of the vehicle types or filtered out This could be due to the sensor array s vehicle classification configuration since all bicycles were detected correctly as motorcycles Also arm and leg movements occasionally caused 10 multiple detections The sensor scan rate probably needs to be reduced and range meas urements increased for proper pedestrian detection Dharmaraju et al 2001 2 4 Passive infrared sensors Hashimoto et al 1997 1998 1999 2000 developed a direction sensitive people detec tion system based on a PIR sensor The sensor contained a one dimensional eight element sensor array made of pyroelectric lead titanate PbT103 ceramics an IR trans parent lens and a motorized chopper Figure 3 The chopper was used to cut the IR ray to the sensor array on and off By do
138. stem for mobile communications Graphical user interface Hue saturation and intensity Hertz Indoor air quality IR ITS kg LVQ lx MICA MIDE MOCUS N A PbTi053 PC PET PIR PoE ppm RBF RJ SIM SOM PAK SVGA USA USB USD UK IX Infrared Intelligent transportation system Kilogram Learning vector guantization lux Meter Multidimensional independent component analysis Multidisciplinary Institute of Digitalisation and Energy Moving Object Counting approach using Ultrasound Sensor networks Not available Lead titanate lead titanium oxide Personal computer Polyethylene terephthalate Passive infrared sensor Power over Ethernet Parts per million in a volume Radial basis function Registered jack Subscriber identity module Self organizing map program package Super video graphics array United States of America Universal serial bus United States dollar United Kingdom VGA VHS XML a Volt Watt Visibility graph analysis Video Home System Extensible markup language Centigrade degree Celsius 1 Introduction 1 1 Background and thesis objectives People flow measurement is a valuable tool in various applications e g in mapping market strategies Adriano et al 2005 For instance at shopping malls one of the most important attributes determining the value of individual stores is the number of people passing it by In fact the success or failure of a shopping center as
139. t The mat sensor s counting accuracy is thus generally better for higher visitor flows compared to an IR beam sensor whose readings are usually compensated for overcount ing with a certain delay between adjacent pulses Teknovisio Ltd 2009 2011b 75 um PET 3 layers shield 9 um Al signal electrode 9 um Al Emfit film 65 um ground electrode 9 um Al Figure 29 Structure of Emfit piezoelectric sensor film left and layout of the sensor elements right Emfit Ltd 2003 The Schmersal SMS 4 switching device mat consists of two separate conductive steel plates held apart by insulating strips The sensor mat is covered in chemical resistant polyurethane Figure 30 left When activated by a step on the mat a short circuit is produced between the conductive plates Figure 30 right The mat is originally de signed for safety purposes to cut power from hazardous machinery in case somebody ventures too close to them Its response time is 25 ms at the most and its limitations in separating simultaneous passers by and recognizing walking directions are similar to the Emfit sensor mat Schmersal s mechanical life is promised to be at least 1 5 million operations and it needs an applied weight of 20 kg to work properly Hence the sensor mat is quite undurable to be used in people flow measurement applications at least for very busy locations or suitable for counting light weighted people like children The Schmersal sensor mat needs
140. t sign visible at the top of Figure 58 in section 5 4 that could not be removed during the tests limited the camera s field of view The Axis video camera was the only sensor using its own data collection software and its counting data had to be downloaded later via an Ethernet connection The clock of the Axis video camera sensor was found out to drift significantly during a couple of days thus causing a dilemma in synchronizing it with the rest of the system While configuring the Irisys IR camera sensor it was found out that the surrounding metallic air ducts caused reflections of the heat emitted by the passing persons leading to double detections Fortunately the problem was fixed by adjusting the counting area of the sensor As a result however the width of the count ing area slightly narrowed The side mounted IR light beam sensors were also placed at slightly different locations in the corridor direction This was due to the fact that they had to be installed at the same height and to acknowledge the notification that the Takex and Cedes sensors if placed too closely interfered with each other The doorway was also wide enough for two or three persons to walk through simultaneously Although in the preliminary ob servations this occurred relatively seldom it is a clear source of error in the counting results of the IR light beam sensors The dimensions of the Schmersal sensor mat used in this test setup were 100 cm x 50 cm The mat wit
141. tal parameters is presented in Figure 56 The settings used for the Eurotech stereoscopic camera were the distance between the sensor front panel and the upper border of the detection area was 31 40 cm the light intensity setting was chosen to be automatic and the door threshold was set to the cen ter of the camera s field of view A sample view of the Eurotech s configuration soft ware is shown in Figure 57 EJ SOPAS Engineering Toot TYS100 Ae Ek Connain Wen Tous Heo ea vu Ja X Show contour 7 Show sagenants 2 Show counter ores Bos ders of countro ores vin a B Cperstce O 1 S100 20000 COMI cnino W sndroneed O Donnlosd Immedstely Figure 56 Configuration software of the Sick time of flight camera sensor The graph illustrates the counting area in red and the contour of the sensor s detection beam in purple and turguoise Correction Combats 120 Sennge Wide Gate Adve tle Date and Tene settinga Date DIRMA Time hh meri osm Sel Ode end Tire Distance Configuration C 22 30 cm 31 40 n Light intensity 19 S amp fonestic In Out direction C int Ont Int Out 20410 Out 19783 Door kind 8 Tyo wae One way lh Door threshold 4 View Thero no D sta record Sawe tacarcs PCN Convigurctan fies masoures Load PCF Razoo Facey Setige Figure 57 Configuration software of the Eurotech stereoscopic camera sensor The two upper left views disp
142. tering of the head pixels into regions or blobs and finally filtering the noise The people counting module screen was divided into three zones top track ing and bottom When a blob passed all three zones continuously a count of certain direction was made Temporary vectors were used to keep a record of blobs coordi nates and zone history Tracking was performed by counting the Euclidean distance between the blobs of successive frames The captured video was sampled in 20 frames per second and saved in AVI format Due to the size and resolution of the captured im ages the system was unable to run in real time The tracking speed however could be further increased by using only the values of alternating pixels Adriano et al 2005 The neural network of APeC s people detection module was trained with sample values of pixels representing people s heads The program however tended to group even in significant noise spots to head regions and dark objects close to each other were identi fied as one region Shadows and luggage were as well easily construed as heads Using size constraint thresholds these problems could be reduced to a certain limit The count ing algorithm was also found to be capable of handling the following exceptional situa tions an object going back and forth between two regions before exiting a missing de tection of a head between two frames of correct detections and a background object being wrongly ide
143. the detected areas in the thermal image Two persons could be detected individually even if they were moving in same or opposite directions The people counting system was 11 evaluated at a door of 100 cm in width and 220 cm in height while altogether 2 152 persons passed freely through the door during an observation period of two hours The sensor s chopping frequency was adjusted to 20 Hz to detect people moving 5 m s The achieved minimum detectable distance between two persons was about 10 cm and the detection accuracy of the movement direction and the number of passers by was 99 6 and 98 5 respectively Among other factors large hand and leg movements caused detection errors Hashimoto et al 1997 processing circuit count sensor computer V moving direction sensing areas Figure 4 People detection system based on a one dimensional eight element array PIR sensor by Hashimoto et al Hashimoto et al 1997 The sensor development by Hashimoto et al continued with a people counting system extended to include three PIR sensors capable of covering a 2 m wide doorway If the people crossed the monitored area between two sensors or across fields of multiple sen sors an accurate detection of the number of people was impossible Thus an extra IR sensor was installed in the middle of the sensor array and perpendicular to the counting direction to detect the width of the passing pattern i e the number of
144. this previous test only pedestri ans without any direction information were registered and the counting results were put on paper in 10 minute intervals The field observer used by Diogenes et al however handled their assignment poorly For example they started counting late they didn t take note of their breaks or they registered bicycles as pedestrians The EcoCounter 14 sensor was found to systematically undercount the pedestrians and had an error rate of 19 9 being 13 2 on average with a standard deviation of 0 14 The occasional overcounting was probably due to the pedestrians wandering around in the detection area as mentioned above The sensor manufacturer had given the device a counting error of about 5 and informed that it might miss adjacent pedestrians Pedestrian vol umes were found to have no notable effect on the sensor performance but pedestrians walking in groups or otherwise close together were found to be easily missed The rela tively stable error of the sensor could be possibly compensated by a general or even site specific correction factor To separate the pedestrians and cyclists the sensor manu facturer had also suggested using the passive IR sensor simultaneously with a pneumat ic tube that would register only bicycles Greene Roesel et al 2008 Mathews amp Poign 2009 designed a portable and reasonably priced sensor usable in public spaces The hardware consisted of a counter of four PIR sensor
145. tom IR photo sensor Cedes 9 TPS 210 Two beam triangulation Yes 150 x 48 x 26 106 200 IR sensor Irisys IRC3020 Thermal array based Yes 111 x 111 x 70 200 220 480 198 452 people counter recommended recommended 250 450 225 425 optimal 350 optimal 325 Eurotech PCN 1001 Stereoscopic camera Yes 100 x 230 x 515 240 1209 44 5 70 Sick TVS 100 Time of flight camera Yes 150 x 180 x 110 5 2300 230 550 100 650 Axis M3203 Network camera Yes 132 x 144 x 94 430 250 500 250 500 recommended recommended 300 500 300 500 a Emfit Ltd 2003 b K A Schmersal GmbH 2010a 2010b c Yu Heng Electric Co Ltd 2005 2008 d Takenaka Electronic Industrial Co Ltd 2006 e Cedes AG 2009a 2009b f InfraRed Integrated Systems Ltd 2009a 2010 g Eurotech S p A 2010 2011 h Sick AG 2009 2011 1 Axis Communications AB 2010a 2010b Cognimatics AB 2010a J The test device was a combination of two sensors a custom circuit and casing assembled by Teknovisio Ltd k The test device was a combination of a sensor mat and a pulse module with casing assembled by Teknovisio Ltd 1 Height x width x depth thickness m Without possible mounting brackets or removable cabling n Maximum installation height if minimum person height is set to 1 m o Multiple sensors can be coupled together to widen the monitored area Brand Model Idle mode output Supply voltage Power source Power Operating
146. tor vehicles are present as well test 2 were included The counting error rates for overall counts and cyclist classifications were less than 10 and less than 5 respectively in test 1 In one experiment a strong wind however in creased the rate of false observations In test 2 the overall counting error was 0 49 and the cyclist classification error 2 16 Possible cars in the camera s field of view mainly caused increment in the counting results There were also difficulties to place the counter in a way that it was able to see both sides of the road which in turn led to undercounting Shadows also caused disjointed trajectories and thus false counting Heikkil amp Silv n 2004 20 Tracking unit CCD camera frame grabber E processor card memory card slot battery hand held g computer for setup EEE Analysis Figure 10 Structure of a video camera based system for outdoor monitoring of cyclists and pedestrians by Heikkil amp Silven Heikkil amp Silv n 2004 In an Automated People Counting from Video APeC application by Adriano et al 2005 a single video camera was placed above a doorway Passing persons were rep resented in the captured video frames by their heads Object detection of the image pro cessing software was divided into three phases pixel classification to head and non head classes using the self organizing map program package SOM_PAK a C pro gram package clus
147. tps www my sick com PDF Create aspx ProductID 36530 amp Culture en US Song J amp Dong Y F amp Yang X W amp Gu J H amp Fan P P 2008 Infrared passenger flow collection system based on RBF neural net In 2008 International Conference on Machine Learning and Cybernetics Kunming China 12 15 7 2008 Piscataway NJ USA The Institute of Electrical and Electronics Engineers Inc pp 1277 1281 ISBN 978 142 44209 57 SRF Consulting Group Inc 2003 Bicycle and Pedestrian Detection Final Report Minneapolis MN USA 29 p Takenaka Electronic Industrial Co Ltd 2006 Takex DL S series background suppres sion photo sensors Datasheet Kyoto Japan 12 p Referred 18 5 2011 Available http www takex elec co jp product suppression pdf Cat_DL S_E pdf Teirikangas J amp Uotinen S amp Kuutti J amp Jokelainen J amp Tuoma H amp Kivilahti A 2010 Anturiverkko mittaa v enkulkua Prosessori Vol 32 5 pp 20 23 ISSN 0357 4121 Teknovisio Ltd 2009 Visit Mat pressure sensitive mat sensor Web document Re ferred 20 12 2011 Available http www visit fi uk visit outdoor mat htm Teknovisio Ltd 2011a Visit Log GSM Local data storage unit with GSM data trans mission Web document Referred 10 10 2011 Available http www visit fi uk visit_ log gsm php Thornton S J amp Bradshaw R P amp McCullagh M J 1991 Pedestrian Flows and Retail Turnover British Food Journal Vol 93 9 p
148. ts The CO sensor based DCV maintained the minimum airflow at 1 s m until the room s CO concentration rose to 900 ppm After this the ventilation was increased and regu lated to keep the CO level at constant 900 ppm After the lecture was over and the CO level dropped below 700 ppm the minimum ventilation rate was used The pre supposed CO production rate was 9 mg s per person and the CO 2 outdoor level was 350 ppmy The IR occupancy sensor used the settings of the CAV and turned the venti lation to the designed maximum when the room was occupied and to the minimum when the room was empty The occupancy sensor s weakness was overventilation when the number of people in the room was less than the maximum This led to a waste of energy and a possible feeling of draught and an uncomfortably low temperature IR sen sors capable of recognizing the number of people in a room could be one possible solu tion to overcome these drawbacks The two main factors affecting possible energy sav ings of DCV systems were the actual occupancy density and the number of hours that the ventilated areas were in use The required fresh air volumes per day with normal school activities were calculated for the three ventilation strategies Additionally infor mation was collected about the number of people assigned to class and the actual num ber present in the classroom the use of the room during inspection and the floor area and volume of the room
149. ttp mide aalto fi en 4D Space Aalto University 2011b Multidisciplinary Institute of Digitalisation and Energy MIDE Web document Referred 17 5 2011 Available http mide aalto fi en MIDE Adriano G P amp Mendoza S I V amp Montinola F N J amp Naval P C 2005 APeC Au tomated People Counting from Video Research document Ouezon City Philippines University of the Philippines Diliman College of Engineering Department of Comput er Science 5 p Amin I J amp Taylor A J amp Junejo F amp Al Habaibeh A amp Parkin R M 2008 Auto mated people counting by using low resolution infrared and visual cameras Measure ment Vol 41 6 pp 589 599 ISSN 0263 2241 Anberger A amp Hinterberger B 2003 Visitor monitoring methods for managing public use pressures in the Danube Floodplains National Park Austria Journal for Nature Conservation Vol 11 4 pp 260 267 ISSN 1617 1381 Axis Communications AB 2008 AXIS PoE Midspan 1 Port Installation Guide Ver 1 00 Lund Sweden 32 p Referred 20 10 2011 Available http www axis com files manuals ig_poe_midspan_32302_en_0805 pdf Axis Communications AB 2010a AXIS M32 Network Camera Series Datasheet Lund Sweden 2 p Referred 14 10 2011 Available http www axis com files datasheet ds_ m32_41247_en_1011_lo pdf Axis Communications AB 2010b AXIS M32 Series Fixed Dome Network Camera Us er s Manual Rev 1 1 Lund Sweden 49 p Referred 18 5
150. unaaaa unu unnaaaaunuuns0 50 4 1 TECHNICAL DETAILS OF THE TEST SENSORS cansi anin e a am omit s ava T posse 50 Aid PRESSURE SENSITIVE SENSOR MAT So cothesstlasiisescedsmsiosageinwtedoneadiades TE 51 VI 4 3 INFRARED LIGHT BEAM SENSORS msn oiva imo ai Maarata EN AS 53 4 4 INFRARED CAMERA AND VIDEO CAMERA SENSORS cccesssscceeeessseeeecessseeecesssaaeeesesssaeeeseesaaes 55 5 SETUP FOR SENSOR COMPARISON TESTS sseessosessocesssoessocesssocessosessoeess 59 Jl TESTS ERE LOCA TION snai E N A 59 ee SENSOR ASSEMBLY sskstnsosammse suytase T tiransanmendabhiemeaaatas 59 J DATA DOGG ING 2asapitss tsteees taste vesi te aaa aa is it t te diet eed Stared Aa 68 5 4 SENSOR CONFIGURATIO N ssasrssaatuommava s ie moestaao kaan aassa eimie asmsta T insat t s uvaste imkteke maesa T2 6 SENSOR TEST DATA COLLECTION cccccccsssessccsssscsecsassecscsacsecsecsssessccssces 75 6 1 ON THEDA TA COLLECTION crec aia A E E A vos 75 02 COUNTING REPORTS Aian A E E iautaplveneepataeaes 76 6 3 PRECISION TEST DATA nonren nnani T mi tatnod mnounbensseascenasauede maneniteuaauehessnerness 78 7 CONCLUSIONS AND FUTURE WORK seooessoocessscccsssoccecsoocesssocecssoceessoscessssoee 79 7 1 CON EU SONS ssc ces E TE N E O ET E ans 79 12 FPOTURE WORK iarta saraqussieeeiuaycanoeaadineoiae aasoqetiaseiusearceanione 82 BIBLIOGRAPHY 5 sta vosse secs sssceesesanssssescsavscesesusssssoscescecssssscesscssbssescsasssssussssesaesusasces 83 Symbols and abb
151. urotech stereoscopic camera sensor left and its detec tion area right Eurotech S p A 2010 The Axis M3203 is a color network camera sensor with SVGA super video graphics array resolution and an embedded Cognimatics True View people counting software Figure 37 In this test setup the camera was installed without the transparent plastic casing visible on the left in Figure 37 The sensor has no pulse outlets and works on a stand alone principle instead connected to an Ethernet connection through an RJ 45 plug Power is supplied to the sensor is through the same connection using the Power over Ethernet PoE standard The sensor can be fed through to a PoE capable network switch but here the Axis PoE midspan 1 Port was used Axis Communications AB 2008 2010a 58 The Axis sensor is capable of counting simultaneously several people moving in sepa rate directions The counting of passing people is done locally on the sensor s pro cessing unit hence reducing the load and requirements of the local area network con nection A computer is needed only when configuring the sensor or accessing the count ing data The counting software can be logged using any internet browser software and the bi directional counting data is available e g in CSV comma separated values and XML extensible markup language raw data formats Some statistical graph creating functions are also available The sensor s internal memory can store up to one mo
152. ut the data collection system was found to be invisible to trail users demanding very little maintenance especially if solar power was used but expensive capital costs Access to counting data was also flexibly available over the Internet to all authorized people Melville amp Ruohonen 2004 2 6 Video and infrared camera sensors Newman et al 2002 conducted a video camera sensor based customer counting exper iment at a major clothing discount retailer in North East England UK Four closed circuit television CCTV cameras were ceiling mounted at the entrance exit checkout area and customer service area One of the devices was a color video camera and three were of a black and white type The cameras were connected to a VHS Video Home System video recorder through a multiplexer The system used a time lapse function and hence provided an image quality of only five frames per second This made the cap tured pedestrian movement jerky and thus hard to analyze for tracking The raw data was digitized and saved in AVI audio video interleave format for further analysis A hybrid object tracking algorithm based both on skin color information and shape of a face was applied to increase the accuracy Newman et al 2002 The traditional motion detection method based on frame difference is only suitable for continuous low speed object movement Thus the algorithm used here tracked pixel blocks from the previous images to various new posit
153. uture routes of individual pedestrians is available a turn probability matrix can be created This would allow a WONKA user to inject pedestrians at any point of the net work and follow their motion through it Thornton et al 1991 Sonnonc gt W O O Q O ju oonnaand gt C o lt poone O v Figure 23 The junction turning problem Thornton et al 1991 42 People flow measurements based modeling can be exploited in planning new pedestrian routes e g in shopping centers or large scale events Modeling outputs can include flow charts on a certain route or level of service graphs Applying a simulation model and changing selected variables like entrances and exits in use predictions about the influences of varied customer behavior can be made However pedestrians may contin uously change their direction and speed and interact with their surroundings and other pedestrians Thus more detail than just starting and end points of the walking routes is usually necessary for modeling Paavola 2010 UCL Depthmap is one example of spatial network analysis software that can be exploit ed to understand social processes in built environments The program can be used at building scales from small urban to entire cities or provinces Topologically based measures like the integration of adjacent spaces and isovists the volume of space vis ible from a certain point in space substantially correlate with pedestr
154. very light work Fifty five students attended a class and the room CO concentration reached a steady state in about 80 100 minutes It was found out that the profile of the number of 45 occupant estimated by the model mentioned above followed the profile of the CO2 con centration In the steady state the estimation closely represented the actual occupant number Leephakpreeda et al 2001 A real time ventilation control using the mathematical model was tested during a three hour class that included two 15 minute breaks The designed ventilation rate used was determined according to the ASHRAE Standard 62 1989R using the room area and the estimated occupancy level To obtain the energy consumption levels the ventilation loads in both the DCV and the original full ventilation cases were calculated according to the same standard Based on the test results the DCV system efficiently reduced the CO concentration close to the full ventilation system and reduced the energy consump tion by 32 Leephakpreeda et al 2001 Yoshiike et al 1999 developed a 360 degree human information sensor that could be attached to the ceiling and was able to count the number of room occupants their loca tions and movements The information gathered with the sensor further could be ex ploited in the automatic control of the environmental settings to increase indoor comfort and to achieve possible energy savings The sensor could also be applied to security
155. w sensor test data collected in this thesis and the results will be exploited in the mutual performance comparison of the test sensors Given the test setup s limited installation heights for the overhead camera sen sors follow up tests will be arranged at a more optimal location In the follow up test manipulated pedestrian patterns with a variable number of people will be used For the most promising sensors a real environment test at a local shopping mall is scheduled as well Additionally the development and testing of people flow sensor applications is planned Within the framework of the MIDE 4D Space project a wider network of people flow sensors will be designed and installed at Aalto Design Factory Suitable sensors will be selected based on the performance tests and the data collection will be realized using wireless data loggers with an option for real time and individual logging of every pedes trian The collected data will be exploited in further studies of public space people flow modeling Also the sensor network s suitability for larger building complexes like shopping centers will be investigated Future work will also include studies on utilizing people flow sensors in the automatic control of environmental conditions like ventila tion sound and lighting 83 Bibliography Aalto University 2011a Multidisciplinary Institute of Digitalisation and Energy 4D Space Web document Referred 17 5 2011 Available h
156. well as the stores wherein is ultimately determined by the circulation of shoppers Brown 1991 With better information about the shopper circulation through customer flow measurement the shopping mall administration can define more accurately correct rental levels for individual tenants Knowledge about the actual distribution of flows inside the shopping center is also beneficial for the tenants and gives them information about the influence of mutual placing of the shops and stores Providing customers a comfortable and cozy shopping environment hence is beneficial for the mall proprietors Knowledge of the people flow sensor performance is of paramount importance The more reliable data a sensor provides the better are the results of applications utilizing it like people flow modeling Cessford et al 2002 The most fundamental and least tech nical registering method of people flow is manual counting by human observers The process 1s however tedious expensive and typically used only for a short period of time which makes the results guite unreliable in general Heikkil amp Silven 2004 Hence manual counting is usually used only to check the accuracy of sensor based counting People flow sensors can also be utilized in so called smart space applications The sen sor data can be used to determine the number of people occupying a certain space and further use the information to adjust the environmental conditions such as ventilation an
157. width or the covering soil material In the operation tests of the sen sors children weighting under 15 kg or people stepping of the border of the sensor were in some cases found to be missed A group of six persons walking freely caused an error of up to 53 in the sensor reading The correct counting of people walking one after the other required at least a one meter separation between consecutive pedestrians Limiting the path width would usually make people keep a sufficient walking distance Trigger ing the counter by stepping on the center of a slab was easy with every tested cover ma terial The softest cover materials like wood chips however passed on the step pres sure rather poorly Ruph Haller et al 2006 Yang et al 2011 tested the counting performance of the EcoCounter dual pyroelectric IR sensor and created a calibration procedure to enhance its counting accuracy The test sensor was equipped with a four threshold algorithm for false count avoidance caused by the movement of rain sun or possible local vegetation The EcoCounter s data log ger is able to store up to one year s data in 15 minute intervals The counting sensor was installed at three sites located in New Jersey USA that were selected based on the pe destrian volume system mounting possibilities accessibility and suggestions by local transport authorities Counting data was collected for 9 hours at site 1 at site 2 in two periods lasting 11 and 12 hours
158. x S ANNNNNNNNNNSISNSN uaaa 151 KAYTAVA w 2 81 wm a VG Yin zi sa 4 Emergency exit E 38 i 15 38 28 18980 1588 Et ag 153 yma PUUTYOSAT N uam 2 PATA g N x ra 156 154 v KUIVAAMO KYLLASTAMO DaS 1519 138 HA ATA AT AT AAT AAT AAW AY AAT A RAN AT bn YB BABA BY BY AASB AAT AA OT OY to OT A A A S lama PO UupOsImEL vun ai KIEN issa v 1 10 sm wm D eae z SIE I F JOE ISZ II 0 ITE p INS PIIPIIIT y LE 88 AAA ANAA en AS sil d i 555500 EI60 38585 Figure 38 Plan of Design Factory s main floor the upper picture continues on the right in the lower picture The test site location is marked with a red circle h denotes height and w width in meters Modified from Parviainen Arkkitehdit Ltd 2003 61 Figure 39 Test site as seen from the corridor Sensors A Emfit B Schmersal C Clas Ohlson D Takex E Cedes H Eurotech and J Axis 62 Figure 40 Test site as seen from the caf Sensors A Emfit B
159. y at higher pedestrian densities Several IR camera sensors from Irisys can be linked together to cover a wider area Bu et al 2007 IR camera sensors can better handle challenging lighting conditions than visual cameras They can be possibly utilized in people tracking in their detection area as well Chan amp Bu 2005 Pulse ultrasound sensors send pulsed waves and measure the propagation time of the reflected echo Continuous ultrasound sensors send the waves continuously and use the Doppler principle for object detection The Doppler principle can be used to determine the object s speed and direction but not to detect stationary objects For ultrasound sen sors a preferred assembly is facing directly downwards or horizontally sideways to the monitored area As the speed of sound varies according to the temperature and medium the pedestrian s clothing and when operating outdoors weather conditions affect the detection Chan amp Bu 2005 The operation principle of microwave radar sensors is similar to ultrasound sensors They send radio waves and detect changes in the waves reflected from the moving ob jects Microwave radar sensors are relatively small and can be set to detect both pedes trians and vehicles They can also feature an adjustable detection interval Cessford et al 2002 An ultra wide band radar sensor sends and receives an extremely short dura tion radio frequency bursts and has good potential in intelligent transpo
160. y provide a handy tool in verifying the proper operation of the sensors they could also tempt people to stop in the sensor area Using the test setup of this thesis only the measurement data of free people flow was collect ed Although the data was collected for a relatively long period and with high resolu tion controlled measurements with predetermined pedestrian patterns should also have been done This would have provided information about the fundamental accuracy and problematics of the tested sensors 81 The test sensors can be roughly grouped to low and high volume devices according to their detection and counting capacity The sensor mats and the side mounted IR beam cells cannot correctly count many persons passing simultaneously and can thus be clas sified as low capacity sensors They are suitable for locations where people are either forced to walk in a single file e g using barriers at security checkpoints or where the pedestrians pass the counting site mostly one by one and a certain level of error caused by occasional groups can be tolerated The overhead mounted camera sensors are capa ble of detecting even people walking side by side Hence they have a high counting capacity and can be utilized at locations of high people flow rate In selection of the people flow sensor also the price is inevitably an important issue The price spread of the test sensors wholesale prices excluding value added taxes was between EUR

A Test Setup for Comparison of People Flow Sensors

Contents

Download Pdf Manuals

Related Search

Related Contents