www.redpel.com +917620593389

Real Time Building Zone Occupancy Detection and Activity Visualization Utilizing a Visitor Counting Sensor Network Jussi Kuutti 1, Petri Saarikko 2 and Raimo E. Sepponen 1 1

2

Health Factory, Aalto University, Espoo, Finland Department of Media Technology, Aalto University, Espoo, Finland

Abstract—Demand-controlled ventilation (DCV) can be adjusted based on room occupancy levels. In this study the feasibility of a visitor counting sensor network in occupancy detection was evaluated. A network with 15 sensor spots and real time activity visualization was designed and assembled at the Aalto Design Factory building. Direction sensitive light beam and infrared (IR) camera sensors were used. Counting data was collected for one week. The sensor spots divided the building into ten zones and the zones’ occupancies were calculated in five minute intervals. The results suggest that visitor counting errors easily accumulate over time and the use of correction factors or a more sophisticated counting algorithm is needed. The operation of a visitor sensor-based DCV could also be complemented with CO2 sensors to guarantee both sufficient ventilation and a short response time. Index Terms—Visitor counting sensor, occupancy detection, demand-controlled ventilation, wireless sensor network, real time visitor monitoring

I. INTRODUCTION Visitor counting data can be exploited in the determination of room or building occupancy levels for the adjustment of demand-controlled ventilation (DCV) [1,2]. Ventilation recommendations usually determine the required outdoor airflow rates as units of volume per person [3], and during unoccupied periods the ventilation system can therefore be turned down to minimum or shut off completely. Hence, through the use of DCV, the energy consumption compared to a constant airflow ventilation system can be reduced without impairing the indoor air quality [3,4,5]. A problem with the currently widely used CO2 sensor-based DCV is a long response time, resulting from the time it takes for the CO2 level to build following an increase in the room’s population [6]. On the contrary, a visitor counting system responds to change in the room’s population essentially without delay. A visitor counting system also provides information about the movement of people around different parts of a building, which can be used for commercial or security applications [7,8]. In previous research Berger and Armitage (2010) used low resolution infrared (IR) cameras and Wahl et al. (2012) used passive infrared (PIR) sensors for room occupancy measurements [9,10]. Erickson and Cerpa (2010) and Hutchins et al. (2007) described different models for predicting building occupancy levels based on people counting sensor data [6,11]. Meyn et al. (2009) introduced a system for estimating building occupancies

based on data from a sensor network including PIR and video camera sensors and CO2 detectors [7]. Lin et al. (2011) made a study on exploiting irregularities of radio signals in existing wireless networks to detect people present in a space [12]. Melfi et al. (2011) proposed that changes in a building’s information network activity and electricity consumption could be used for implicit occupancy sensing [13]. The objective of this study was to implement a network of 15 visitor counting sensors that could be used to collect counting data at selected locations of the test building. The collected data was used to calculate the occupancy levels as a function of time in the zones bounded by the sensors. The system’s applicability for controlling a DCV was evaluated based on the calculated zone occupancies. II.

MATERIALS AND METHODS

A. Visitor Counting Sensors The test setup included twelve wall-assembled light beam sensors and three ceiling mounted IR camera sensors (Fig. 1). Both sensor types were direction sensitive and provided a counting pulse outlet for each counting direction. The operation of the active two-beam triangulation IR sensor (TPS 210, Cedes, Landquart, Switzerland) is based on measuring the angle of the received reflected light [14]. The sensor is side-mounted and needs no separate reflector plate and its direction discrimination is based on the cutting order of the two IR beams. The sensor’s sensing range can be adjusted between 0.3 m and 2 m, and the maximum speed of direction recognition for a 0.2 m wide object is 5 m/s [15]. The light beam sensor cannot detect multiple people walking in parallel or too close to each other successively. Thus compact groups of people will be inevitably undercounted. The duration of one counting pulse (200 ms) also limits the discrimination between pedestrians walking closely after each other [15]. For the light beam sensor in question, the counting accuracy of one hour periods with free people flow has been as high as 95% in evaluation tests performed previously by the authors. The visitor detection of the thermal array based people counter (Irisys IRC3020, InfraRed Integrated Systems, Northampton, UK) is based on processing the IR image captured by the sensor. The counting software of the sensor shows total counts plus both a video view of the counting site and the IR camera’s detection visualization. The recommended installation height of the sensor is 220–

978-1-4799-2024-2/14/$31.00 ©2014 IEEE Polytechnic of Porto (ISEP) in Porto, Portugal from 26-28 February12014 2014 11th International Conference on Remote Engineering and Virtual Instrumentation (REV) Page 219

www.redpel.com +9176205933891

480 cm with a 60 ° field of view, being suitable for a passage width up to 4.2 m. The sensor is capable of detecting multiple people simultaneously and its counting accuracy is mentioned to be as high as 98% [16,17]. B. Visitor Counting Data Loggers Each visitor counting sensor was equipped with a local data logger that was built around an Atmel ATmega328Pbased development board (Arduino Uno R3, Smart Projects, Scarmagno, Italy) [18]. A visitor counting sensor was connected to two of the development board’s digital input pins featuring hardware interrupts. A data logging shield for the development board (Data logging shield for Arduino v1.0, Adafruit, New York, NY, USA) with a DS1307 real time clock (RTC) connected via the I²C bus was connected to the board with pin headers [19]. The logging shield was equipped with a 4-GB SDHC memory card (TS4GSDHC6, Transcend, Taipei, Taiwan) for storing of counting data. One visitor sensor and data logger pair shared the same 12 V DC power supply. The data loggers were connected to a server computer through two ZigBee gateways. Data loggers were of two types (Fig. 2): five of them were equipped with a 1.25 mW ZigBee radio module (XBee ZB, Digi International, Minnetonka, MN, USA) and ten of them featured a 63 mW ZigBee radio module (XBee-Pro ZB, Digi International, Minnetonka, MN, USA) for a longer range [20]. The radio modules were connected to the development boards using ZigBee adapters (XBee Adapter kit v1.1, Adafruit, New York, NY, USA) [21].

Figure 1. The light beam (left) and IR camera (right) visitor counting sensors used in the test setup.

Figure 2. Two versions of the visitor counting data loggers: with a 1.25 mW (left) and with a 63 mW (right) ZigBee radio.

The counting software was programmed using the Arduino integrated development environment (IDE) version 0022 [22]. Some of the software libraries of the IDE were modified to meet the required hardware configurations. Upon a pedestrian pass a counting pulse from a visitor sensor triggers the interrupt handler routine of the data logger’s software. The interrupt handler sets a flag variable indicating the digital input that has been triggered (the walking direction). The software’s main loop checks the flag, retrieves the current time from the RTC, and appends these as a text line in UNIX epoch format to the CSV file on the memory card. The information is also sent to one of the ZigBee gateways that forwards it to a server computer via an Ethernet connection. After this the main loop clears the flag. C. ZigBee Gateways Each of the data loggers was wirelessly communicating with one of the two ZigBee gateways. The first of them (G1) was for connecting with the loggers featuring the 1.25 mW ZigBee module. It consisted of an Atmel ATmega328P-based development board (Arduino Uno R3, Smart Projects, Scarmagno, Italy), a ZigBee adapter (XBee Adapter kit v1.1, Adafruit, New York, NY, USA), a ZigBee radio module (XBee ZB, Digi International, Minnetonka, MN, USA) and an Ethernet shield (Arduino Ethernet Shield R3, Smart Projects, Scarmagno, Italy) [18,20,21,23]. This gateway’s firmware was created using the Arduino IDE 0022 [22]. The second gateway (G2), for the loggers with the 63 mW ZigBee module, was a commercial device (ConnectPort X2, Digi International, Minnetonka, MN, USA) [24]. The gateways receive the time stamp and direction information of the pedestrian passes from the data loggers and forward them immediately to the server computer through an Ethernet connection. The ZigBee gateways are powered with 12 V DC power supplies. A schematic diagram of the complete visitor counting network is presented in Fig. 3. D. Cost Estimates The wholesale prices of the light beam sensor and the IR camera sensor used in the test setup were 150 EUR and 500 EUR, respectively. The component cost for one visitor counting data logger was about 110 EUR. Thus, the cost of the measurement equipment in one sensor spot was about 260 EUR for the light beam sensors and about 610 EUR for the IR camera sensors. The gateways cost about 90 EUR each. The prices exclude value added taxes (VAT). The costs of individual devices and components would naturally be reduced if the purchase batches were of larger scale. E. Sensor Spot and Gateway Locations and Occupancy Zones The sensor network was built at the Aalto Design Factory building in Espoo, Finland. Four of the visitor counting sensors were installed at the exits of the building, four at inside staircases and seven at intersections of rooms and spaces. Each sensor was accompanied by a data logger. The installation height of the light beam sensors was about 122 cm, which was consider to be optimal to prevent counting error caused by swinging arms. People shorter than this (for example children) were expected to rarely visit the building. The IR camera sensors were installed at places where higher counting

978-1-4799-2024-2/14/$31.00 ©2014 IEEE Polytechnic of Porto (ISEP) in Porto, Portugal from 26-28 February22014 2014 11th International Conference on Remote Engineering and Virtual Instrumentation (REV) Page 220

www.redpel.com +9176205933892

Server computer 8

4 3

9

7 10

5 6

ZigBee Ethernet

2

11

Ethernet

Gateway G1

1

ZigBee

Gateway G2

15

Real time activity visualization

14

12 13

Figure 3. Schematic diagram of the building zone occupancy detection and real time activity visualization system. Sensor spots 1, 6 and 12 featured an IR camera sensor and the rest of the sensor spots a light beam sensor.

capacity was likely to be needed and the installation height was sufficient for a camera sensor. The IR camera sensors’ installation heights varied within the recommended limits depending on the installation location. Installation examples of the sensors and data loggers are presented in Fig. 4. The sensor spots divided the building in ten occupancy zones: the whole building and zones A–I inside the building. The locations of the sensor spots, gateways and in-house occupancy zones are presented in Fig. 5.

G. Data Collection Periods Visitor sensor counting data was collected at each sensor location for one week from Thursday 10.5.2013 0:00 to Thursday 16.5.2013 0:00. For each pedestrian pass a time stamp and the walking direction was registered.

F. Real Time Activity Visualization The server computer receives the time stamp and direction of each pedestrian pass from the ZigBee gateways. The information is stored on the computer and immediately visualized over the building floor plan (Fig. 6). The sensor locations are marked in the plan with dots. Upon a visitor pass an arrow indicating the walking direction is shown. The cumulative in and out readings over a predetermined period and the current occupancy level of the building are also shown. The visualization can be utilized e.g. in monitoring people’s movement in buildings for security or administrative purposes. L S

S

L

Figure 4. Installation examples of the visitor counting sensors (S) and data loggers (L) at the sensor spots: with a light beam sensor (left) and with an IR camera sensor (right).

Figure 5. Placement of the sensor spots (1–15), gateway terminals (G1 and G2), and defined in-house occupancy zones (A–I) at the Aalto Design Factory. Sensor spots 1, 6 and 12 featured an IR camera sensor and the rest of the sensor spots a light beam sensor. The sensor spots 1–5 communicated with the gateway G1 and the spots 6–15 with the gateway G2. The counting directions of the sensors are marked with arrows.

978-1-4799-2024-2/14/$31.00 ©2014 IEEE Polytechnic of Porto (ISEP) in Porto, Portugal from 26-28 February32014 2014 11th International Conference on Remote Engineering and Virtual Instrumentation (REV) Page 221

www.redpel.com +9176205933893

B

A

Figure 6. Real time visualization of the visitor detection (main floor of the Aalto Design Factory). The white dots mark the sensor locations and an arrow indicates a visitor pass and its direction (A). The application also shows the building-level cumulative in and out readings over a predetermined period and the current building occupancy (B).

H. Data Analysis For this study the sensor counting data was collected from the memory cards of the data loggers. This was done to eliminate the counting errors caused by data loss due to possible ZigBee connection failures. The occupancies of each zone were calculated in five-minute intervals. For this, the directional readings of every sensor limiting the zone were totaled: the readings with a direction into the zone with a plus sign and the readings with a direction out from the zone with a minus sign. The zone occupancy level determination is prone to cumulative error caused by differences in a sensor’s counting accuracies for inbound and outbound traffic and further, in the case of multiple sensors monitoring the same zone, differences between individual sensors. The error tends to accumulate over time, thus causing falsely increasing or decreasing trend in the zone occupancy. The calculated occupancy might, for example, obtain negative or obviously overgrown values. To compensate for the cumulative errors in the calculation of the zone occupancies, two simple constraints were applied on the results: possible negative occupancies were considered as a random error and were zeroed, and all zone occupancies were reset every midnight when the building was assumed to be unoccupied [6]. Both the original and corrected occupancies of each zone were plotted as a function of time. III. RESULTS AND DISCUSSION The original calculated and corrected occupancies of three example zones—the whole building, Zone A and Zone I—are plotted in Fig. 7. The whole-building occupancies were calculated based on the counting results of sensors 1, 7, 8 and 14, which were situated at the two main floor and two basement level entrances. The results for Zone A were calculated based on sensors 2 and 6, and the results for Zone I from sensor 15 alone. Recall Fig. 5 for sensor locations. As the zone occupancy level should always be positive, it can be deduced that the occupancy readings of the

whole building are clearly distorted by a net outbound overcounting, and the readings of Zone A by a net inbound overcounting (Fig. 7). The occupancy readings of Zone I in turn seem to be affected by periodic net outbound and inbound overcounting. The reason for these distortions can be directional over- or undercounting of the sensors, or both. In the case of the whole building the simple correction added to the results quite satisfactorily preserves the shape of the original graph. However, when applied to the readings of Zone A and Zone I, the nightly zeroing of the occupancies causes clear sharp drops to the graph. The necessary delay during and after each counting pulse limited the light beam sensor’s counting capacity. For the IR camera sensor this was not an issue as its counting program was able to buffer rapid counting pulses. The light beam sensors were also unable to detect multiple persons simultaneously, which certainly caused some undercounting when several people passed these sensors at the same time. Another thing that possibly caused over- or undercounting is the fact that a light beam sensor, if installed too high, misses the shortest pedestrians, whereas at too low a position it tends to register the widely swinging arms of some of the passersby. Visitors moving too close to each other are also a source of undercounting for the IR camera sensors, possibly because their thermal images cannot be resolved as separate and instead appear as one continuous cluster. Usually IR camera sensors tolerate more fluctuations in the ambient conditions than video camera sensors that operate with visible light. However, possible airflows caused by temperature gradients can cause false detections in an IR camera. This could have been a problem for the sensor installed at the one exit of the test building (sensor spot 1 in Fig. 5). It is also possible that even the data loggers missed some of the counting pulses thus causing error in the results. Although the counting data was here examined with a five-minute resolution, the occupancy detection system could also be used to adjust a DCV system in real time. However, unless corrected, the errors in counting would

978-1-4799-2024-2/14/$31.00 ©2014 IEEE Polytechnic of Porto (ISEP) in Porto, Portugal from 26-28 February42014 2014 11th International Conference on Remote Engineering and Virtual Instrumentation (REV) Page 222

www.redpel.com +9176205933894

lead to an incorrect determination of the zone occupancy and consequently an incorrect ventilation rate. The cumulative error in the occupancy calculations can to some extent be compensated with the simple zeroing correction used in this study. However, better reliability could be reached by using counting site specific correction factors for the sensors or by introducing an occupancy detection algorithm that exploits existing statistical occupancy information in addition to the real-time counting results. Using only imaging-based IR or video camera sensors—which are supposed to be more accurate than light beam sensors—the counting error and thus the cumulative error would probably be reduced. However, this would cause a notable increase in hardware costs. It should also be noted that some locations are not suitable for the installation of a down-looking overhead camera sensor. A camera sensor mounted at an angle might in turn be confused by the images of passers-by overlapping in perspective, which could hinder the counting accuracy. The counting error of the visitor sensors chosen has to be taken into account when the minimum ventilation level of the DCV-serviced zone is decided and the threshold for a ventilation set point change is determined. The operation of a visitor sensor-based DCV could also be complemented with another sensor technology, like CO2 sensors, to always guarantee both sufficient ventilation and a short response time of the control. This, however, would inevitably increase the cost and complexity of the control system.

IV. CONCLUSIONS Directional visitor counting can be used to control DCV in real time. The test results, however, suggest that visitor counting errors easily accumulate over time and use of site-specific correction factors or a more sophisticated counting algorithm is needed. The operation of a visitor sensor-based DCV could also be complemented e.g. with CO2 sensors to guarantee sufficient ventilation along with a short response time. The sensor-logger pairs could be further equipped with a rechargeable battery to make the sensor network completely wireless. ACKNOWLEDGMENT The research work described herein was completed as a part of the 4D-Space and Hybrid Sense projects of the Aalto University MIDE research program. The visitor counting sensors used in the study were lent by Teknovisio Ltd. The visitor counting data logging system was assembled and programmed by Niklas Ekman from the Aalto University Department of Media Technology. The Aalto Design Factory building in Espoo, Finland served as a test site for the sensor network. The Design Factory personnel provided the authors help with local area network connections and other practical issues. The complete text was proofread by Juhana Leiwo from the Aalto University Health Factory. REFERENCES [1]

Zone occupancy (Number of people)

150 100 50 0

0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0

-50 Time (a)

Zone occupancy (Number of people)

50 40 30 20 10 0

0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0

-10 Time (b)

Zone occupancy (Number of people)

20 10 0 -10

0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0

-20 Time (c) Original data

Corrected data

Figure 7. Occupancies of selected zones of the Aalto Design Factory in five-minute resolution: whole building (a), Zone A (b) and Zone I (c). The results based on the original sensor data, and results corrected by zeroing the negative occupancy values and resetting the occupancy readings at midnight, are presented.

K. Hashimoto, M. Yoshinomoto, S. Matsueda, K. Morinaka, and N. Yoshiike, “Development of people-counting system with human-information sensor using multi-element pyroelectric infrared array detector,” Sensors and Actuators A: Physical, vol. 58, pp. 165–171, February 1997. [2] M. Mysen, S. Berntsen, P. Nafstad, and P.G. Schild, “Occupancy density and benefits of demand-controlled ventilation in Norwegian primary schools,” Energy and Buildings, vol. 37, pp. 1234–1240, December 2005. [3] S.J. Emmerich and A.K. Persily, “Literature review on CO2-based demand-controlled ventilation,” ASHRAE Transactions, vol. 103, pp. 229–243, 1997. [4] K. Hashimoto, C. Kawaguchi, S. Matsueda, K. Morinaka, and N. Yoshiike, “People-counting system using multisensing application,” Sensors and Actuators A: Physical, vol. 66, pp. 50– 55, April 1998. [5] N. Yoshiike, K. Morinaka, K. Hashimoto, M. Kawaguri, and S. Tanaka, “360° direction type human information sensor,” Sensors and Actuators A: Physical, vol. 77, pp. 199–208, November 1999. [6] V.L. Erickson and A.E. Cerpa, “Occupancy Based Demand Response HVAC Control Strategy,” in the 2nd ACM Workshop on Embedded Sensing Systems for Energy-Efficiency in Building, Zurich, Switzerland, 3–5.11.2010, pp. 7–12. [7] S. Meyn, A. Surana, Y. Lin, S.M. Oggianu, S. Narayanan, and T.A. Frewen, ”A Sensor-Utility-Network Method for Estimation of Occupancy in Buildings,” in the Joint 48th IEEE Conference on Decision and Control and 28th Chinese Control Conference, Shanghai, P.R. China, 16–18.12.2009, pp. 1494–1500. [8] S.K. Hui, P.S. Fader, and E.T. Bradlow, “Path Data in Marketing: An Integrative Framework and Prospectus for Model Building,” Marketing Science, vol. 28, pp. 320–335, March–April 2009. [9] M. Berger and A. Armitage, “Room Occupancy Measurement Using Low-Resolution Infrared Cameras,” in the 21st IET Irish Signals and Systems Conference, Cork, Ireland, 23–24.6.2010, pp. 249–254. [10] F. Wahl, M. Milenkovic, and O. Amft, “A Distributed PIR-based Approach for Estimating People Count in Office Environments,” in the 15th IEEE International Conference on Computational Science and Engineering, Paphos, Cyprus, 5-7.12.2012, pp. 640– 647.

978-1-4799-2024-2/14/$31.00 ©2014 IEEE Polytechnic of Porto (ISEP) in Porto, Portugal from 26-28 February52014 2014 11th International Conference on Remote Engineering and Virtual Instrumentation (REV) Page 223

www.redpel.com +9176205933895

[11] J. Hutchins, A. Ihler, and P. Smyth, “Modeling Count Data from Multiple Sensors: A Building Occupancy Model,” in the 2nd IEEE International Workshop on Computational Advances in MultiSensor Adaptive Processing, Saint Thomas, U.S. Virgin Islands, USA, 12–14.12.2007, pp. 241–244. [12] W-C. Lin, W.K.G. Seah, and W. Li, “Exploiting Radio Irregularity in the Internet of Things for Automated People Counting,” in the 22nd IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, Toronto, ON, Canada, 1114.9.2011, pp. 1015–1019. [13] R. Melfi, B. Rosenblum, B. Nordman, K. Christensen, “Measuring Building Occupancy Using Existing Network Infrastructure,” in the 2011 International Green Computing Conference and Workshops, Orlando, FL, USA, 25-28.7.2011, 8 p. [14] TPS – Triangulation Proximity Switch. Available online: http://www.cedes.com/english/Products/DoorsGates/TPS_en.htm (accessed on 21.10.2013). [15] TPS product family. Landquart, Switzerland: Cedes AG, 2010. Available online: http://www.cedes.com/documents/produkte/dg/ 001133en_TPS.pdf (accessed on 21.10.2013). [16] IPU 40421 People Counter Mounting Height Graph. Northampton, United Kingdom: InfraRed Integrated Systems Ltd., 2011. Available online: http://www.peoplecounting.co.uk/files/ IPU_40421_IRC3000_Series_Mounting_Height_Graph_Inches_Is sue_1.pdf (accessed on 21.10.2013). [17] IRC 3020 Relay Counter (60° Master). Available online: http://www.irisys.co.uk/people-counting/irc3020 (accessed on 21.10.2013). [18] Arduino Uno. Available online: http://arduino.cc/en/Main/ arduinoBoardUno (accessed on 21.10. 2013).

[19] Adafruit Data logging shield for Arduino - v1.0. Available online: https://www.adafruit.com/products/243 (accessed on 21.10.2013). [20] XBee & XBee-PRO ZB. Minnetonka, MN, USA: Digi International Inc., 2011. Available online: http://www.digi.com/pdf/ds_ xbeezbmodules.pdf (accessed on 21.10.2013). [21] XBee Adapter kit - v1.1. Available online: https://www.adafruit. com/products/126 (accessed on 21.10.2013). [22] Arduino Software. Available online: http://code.google.com/p/ arduino/ (accessed on 29.11.2013) [23] Arduino Ethernet Shield. Available online: http://arduino.cc/en/ Main/ArduinoEthernetShield (accessed on 21.10.2013). [24] ConnectPort X2. Available online: https://www.sparkfun.com/ products/10569 (accessed on 21.10.2013).

AUTHORS J. Kuutti is with the Health Factory, Aalto University, PO Box 13340, FI-00076, AALTO, Finland (e-mail: [email protected]). P. Saarikko is with the Department of Media Technology, Aalto University, PO Box 15500, FI-00076, AALTO, Finland (e-mail: [email protected]). R. E. Sepponen is with the Health Factory, Aalto University, PO Box 13340, FI-00076, AALTO, Finland (e-mail: [email protected]). This work was supported by Aalto University Multidisciplinary Institute of Digitalisation and Energy (MIDE, http://mide.aalto.fi/en/).

978-1-4799-2024-2/14/$31.00 ©2014 IEEE Polytechnic of Porto (ISEP) in Porto, Portugal from 26-28 February62014 2014 11th International Conference on Remote Engineering and Virtual Instrumentation (REV) Page 224

www.redpel.com +9176205933896