A Wireless Sensor Network Approach to Signalized Left Turn Assist at Intersections

Transcription

1 A Wireless Sensor Network Approach to Signalized Left Turn Assist at Intersections Fabien Chraim, Thomas Watteyne, Ali Ganji,KrisPister BSAC, University of California, Berkeley, USA Civil and Environmental Engineering, University of California, Berkeley, USA, Abstract This paper presents one possible approach to Signalized Left Turn Assist at intersections. A sensing platform is designed to detect the presence of cars by measuring the deflection of the magnetic field when driving by. This platform is tested experimentally under different conditions. A threshold detection algorithm is designed to determine whether a car is passing in front of the sensor. The algorithm is first run off-board using a set of previously collected magnetometer measurements, before being implemented and run on-board the platform. We show how this detection algorithm performs better as the sensor is placed close to the traffic flow and we discuss how such sensors can be networked together to form a smart intersection. I. INTRODUCTION Intersections are a transportation facility which present a major threat to drivers, bicyclists and pedestrians. In 1998, 1.72 million crossing path crashes happened in the USA only, accounting for 27.3% of all police-reported ones in that country. Of those crashes, almost 80% took place at intersections. Yearly, intersection crashes amount to 9000 fatalities (25% of all traffic fatalities in the USA) and 1.5 million injuries (50% of all traffic injuries) [16]. The United States Department Of Transportation (DOT) started a program to develop Intersection Decision Systems to tackle some of the problems that occur on those facilities. Efforts are taking place at the university of Virginia Tech and the University of Minnesota to address both stop sign/light violations and crossing/entering high-speed rural roads from minor segments, respectively. Researchers at the University of California at Berkeley, in coordination with the California Partners for Advanced Transit and Highways (PATH) and Caltrans, are investigating Signalized Left Turn Assist (SLTA) at intersections [1]. Left Turn Across Path/Opposing Direction (LTAP/OD) crashes occur when a subject vehicle (SV) approaches the intersection with the intent to make a left turn through a gap in the stream of principal other vehicles (POV) [2]. One possible solution is to detect the vehicles at or approaching the intersection, and estimate whether the left-turning driver has enough time to make that turn. Such an Intersection Decision Support (IDS) system is an addition to the existing infrastructure seen at signalized intersections. As early as 1962 [17], researchers proposed the use of magnetometers to sense the presence of vehicles. The idea grew quickly in the following decade as highways in the United States started Fig. 1. Typical left-turn scenario. getting instrumented [14], [15]. Recently, magnetometers have been proposed as sensors for applications involving smart parking spaces and vehicle control guidance [9], [11]. With the advent of Wireless Sensor Networks, a simplification tothe transportation applications appeared, making those solutions more practical [11], [12] Recent advances in Micro-Electro-Mechanical System (MEMS) have enabled low-cost, low-power and low-mass 3- axis accelerometers, which can be integrated within a batterypowered sensor node. In this paper, we design and test such a board in various experimental settings, proving that MEMSbased magnetometers can be used for vehicle detection. We design an algorithm to automate the detection. We first run this algorithm on a computer (off-board the sensor) on a set of collected magnetometer traces and prove its efficiency. We then implement it on a typical sensor micro-controller, proving its simplicity and practicality. The remainder of this paper is organized as follows. Section II presents the proposed framework to tackle the problem of LTAP/OD crashes. In Section III, we introduce the sensor platform that we use in our experiments and discuss the vehicle detection algorithm developed for this application, which we test experimentally in Section VI. Section VII takes a look at the system overview from a networking perspective. Finally, Section VIII concludes this paper /11/$ IEEE

2 MSP430 micro-controller (back side) battery connector amplitude of magnetometer car stops antenna time (s) car stopping AT86RF231 radio (back side) HMC5843 magnetometer amplitude of magnetometer car passes Fig. 2. The hardware platform used in this paper time (s) car passing at 40mph II. INTERSECTION DECISION SUPPORT SYSTEM Researchers at UC Berkeley and PATH have proposed a solution to LTAP/OD crashes. As seen in Fig. 1, the subject vehicle A is waiting at the intersection to make a left turn. As the driver is trying to estimate whether it is safe to make the turn, principal other vehicles C and D are approaching from the opposite direction. Ideally, the IDS system can sense vehicles A, B and C, assess the gap in the oncoming stream and flash a no left turn sign if the gap is too small, or if conditions are unsafe (e.g. crossing pedestrians). The warning could possibly also occur inside the vehicle if it is equipped with the appropriate display device. While studying this problem, [1], [2] does not give a final recommendation on the best solution. Rather, it presents a comparison of possible sensor technologies and gives a generic decision algorithm that, based on sensor input, recommends whether or not conditions are safe to make a left turn. III. SENSING PLATFORM We have developed the platform depicted in Fig. 2, to detect passing vehicle, and report the information back to a central unit over a low-power wireless link [3]. This 1.6g sensor board is battery-powered, and features a Honeywell HMC5843 three-axis digital magnetometer [4]. It uses the AT86RF compliant low-power radio by Atmel to transmit data wirelessly [5]. The magnetometer consumes 1mA when measuring; the radio 14mA when transmitting at 0dBm. When requested, the magnetometer outputs three 12-bit values over a digital I2C bus, representing the magnitude of the magnetic field in each of the 3 axis. These are signed integer numbers, where 970 counts represent one Gauss. The Earth s magnetic field has an amplitude of 0.5Gauss; we calibrate the magnetometer for a range of +/- 2Gauss on each axis. Section IV presents preliminary testing of the sensor board. Section V introduces the detection algorithm we developed for this application, and shows results gathered from running the algorithm off-board on previously collected magnetometer Fig. 3. Variation of the amplitude of the magnetometer vector over time; data gathered for the preliminary measurements. data. We then implement on our platform, implementation details are presented in Section V-C. IV. PRELIMINARY MEASUREMENTS The experimental setup for these preliminary measurements involves a single-hop wireless link between the sensor board and a basestation node connected to a computer. The sensor board is positioned onto the pavement, and reports the raw 3-axis magnetometer data every 2.5ms to the basestation over the wireless link. To see the impact of cars on the magnetic field, we have a car stop next to the sensor as a first experiment, and drive the car past the sensor at 40mph. Fig.3depictsthe variation of the magnitude of the magnetometer vector with time. Fig. 3 indicates that the passing of a car can be detected visually. Section V details the filtering and decision algorithm we use to detect vehicles from the magnetometer data. V. DETECTION ALGORITHM AND OFF-BOARD EXECUTION We design a detection algorithm which, given the magnetometer input, detects whether a vehicle is present next to the sensor. Fig. 4 represents the block diagram of the algorithm. In this section, we describe how this algorithm works, and the presents its results when run off-board on previously gathered magnetometer data. In Section VI, we present how we have implemented the algorithm on the sensor board. A. Overview of the Algorithm Using the same firmware as used for the preliminary measurements (Section IV), every 125ms, the microcontroller reads the three axis of the magnetometer and transmits that data over a wireless link to the basestation. The computer connected to the basestation runs the following algorithm, emulating the micro-controller in Fig. 4. The detection algorithm consists of four steps:

3 Fig. 4. Block diagram of the car detection algorithm. ThedatafromthethreeaxisX, Y and Z are combined, and the norm of the resulting vector is calculated as sqrt(x 2 + Y 2 + Z 2 ); A first order Butterworth high-pass filter with a corner frequency of 0.5Hz is used to remove the constant component of the vector (the Earth s magnetic field); this removes the constant offset of the data; Assuming no more than 5 cars pass in front of the sensor per second, we apply a ninth order Butterworth low-pass filter with a corner frequency of 5Hz; The resulting filtered data is fed into the Finite State Machine (FSM) depicted in Fig. 5 and described below. B. Finite State Machine The resulting filtered data is fed into the Finite State Machine (FSM) depicted in Fig. 5. The FSM contains three states: no_car, interim and car. Theinterim state is used as a hysteresis state to avoid oscillations between no_car and car. The FSM has three compile-time parameters mincount, maxcount, threshold (which are not changed) and one variable counter. The FSM starts in the no_car state, counter is initialized to zero. If the filtered magnitude of the magnetometer (called filtered) is lower than threshold, the FSM concludes that there are no cars in the vicinity of the sensor. If the input exceeds threshold, the FSM transitions to the interim state. While in that state, counter in incremented by 1 each time the filtered magnitude of the magnetometer exceeds threshold, and decremented by 1 when it does not. If counter reaches mincount (a negative number), the FSM transitions back to no_car. If it reaches maxcount (a positive number), the FSM transitions to the car state. If, while in the car state, the magnitude drops below threshold, the FSM transitions to the interim state, and counter is reset. A packet is sent every time the FSM transitions from the interim to the car state (indicating a car is in front of the sensor) and when it transitions from the interim to the no_car state (indicating no car is sensed). Conceptually, the algorithm draws a sphere in the 3-D space defined by the axes of the magnetometer, as depicted in Fig. 6. This sphere is centered at a point shifted from the origin by the mean of the arrays composed by each of the three axes (representing the Earth s magnetic field). The radius of this sphere is the threshold used in the state machine. Fig. 6 shows the threshold sphere, as well as individual magnetometer readings. Successive readings which are inside the sphere will cause the FSM to switch to the no_car state; successive Fig. 6. The axis represent the three axis of the magnetometer; the radius of the sphere represents the threshold of the algorithm. Dots represent magnetometer measurements. readings which fall outside of sphere will cause it to transition to the car state. We initialize threshold relative to the standard deviation of the magnetometer norm from its average, and modulate this number by the distance between the sensor and the vehicles. C. Off-board Execution The detection algorithm is tested by deploying three sensor boards on a stretch of road, all running the firmware used for the preliminary measurements (Section IV). As depicted in Fig. 1, we deploy the sensors on a line perpendicular to the traffic, at 1.7m, 2.0m and 2.3m of the main flow of cars. We record the passing of vehicles by hand to establish a baseline truth, and run the algorithm on the gathered magnetometer data on a computer running Matlab, using the flow of magnetometer data received from the sensors. Results over a measurements period of 300s is presented in Fig. 7. Each plot contains three graphs. The bottom graphs represents the hand recordings of the cars passing by (the truth), where high indicates there is a car in front of the sensor. The uppermost graph is the raw magnitude of the magnetometer vector. The middle graph is the output of the detection algorithm described in Section V-A. With 226 cars recorded, we draw table I. 50% of false positives match the passing of a large vehicle (bus or truck) on the opposite side of the street where the experiment was run. Looking at the percentage of false negatives as a function of the distance between the sensor and traffic, one can notice that the relationship is linear. Furthermore, the resulting line has a negative intercept alluding to a near zero error rate in the algorithm. In other words, the closer the sensor is placed to the flow of cars (at distances less than 0.5m), the more accurate the algorithm. In an actual implementation, the sensor boards could be embedded into the pavement or inside the lane reflectors. VI. ON-BOARD IMPLEMENTATION DETAILS In a real setup, it is impractical for each sensor node to stream its raw magnetometer data for off-board processing.

4 Fig. 5. Finite State Machine used as the last step in the car detection algorithm (Fig. 4). 2.3m from traffic 2.0m from traffic 1.7m from traffic Fig. 7. Raw magnetometer data (upper), detection algorithm output (middle) and base truth (lower) for sensors placed at different distances from the traffic flow. Distance False False to Traffic Postitive Negative 1.7m 14 (6%) 53 (23%) 2m 10 (4%) 75 (33%) 2.3m 18 (8%) 93 (41%) TABLE I FALSE POSITIVES/NEGATIVES AS A FUNCTION OF DISTANCE TO TRAFFIC. Rather, each node should process the data on-board, and send high-level information about the sensed phenomenon. In our case, we want each node to send a packet when it starts detecting a car, and a packet when it stops detecting it. The complexity of on-board execution lies in the implementation of the band-pass filter shown in Fig. 4, and which needs to be implemented in firmware on the board s micro-controller. The sensor board is driven by an MSP430f bit micro-controller, with 116kB of flash and 8kB of RAM. We implement the band-pass filter in C as a Lattice Wave Digital Filter. The complete firmware running on the board (including band-pass filter and FSM)takes up 1.7kB of code space and 160B of RAM 1. The code structure follows the algorithm described before; a pseudo-code of the implementation in presented in Algorithm 1. We use the hardware multiplier of the micro-controller to compute the norm of the vector, and to do the filtering. On micro-controllers which do not feature a hardware multiplier, the same algorithm can be implemented using Horner s Scheme, which only uses additions and shift operations. The state kept by the algorithm is not the data itself (e.g. the last 100 measurements), but rather the state and parameters of the state machine (e.g. the value of counter). This greatly reduces the memory footprint of the implementation. The execution of the complete algorithm (filtering and traversing the FSM) takes 12ms; we therefore iterate the algorithm at 80Hz, or one loop every 12.5ms. VII. TOWARDS SMART INTERSECTIONS As future work, we consider the intersection as a network of magnetometer sensors equipped with low-power radios deployed throughout an intersection. When a vehicle enters the intersection, a subset of the sensors will detect its presence and report this information to a central intersection management 1 As an online addition to this paper, the firmware used is made publically available at

5 Algorithm 1 Pseudo-code of the on-board implementation of the vehicle detection algorithm. The loop runs at 80Hz, i.e. one iteration every 12.5ms. 1: loop 2: X, Y, Z magnetometer reading 3: norm norm(x, Y, Z) 4: filtered LPF(HPF(norm)) 5: if state = no_car then 6: if filtered threshold then 7: counter 0 8: state interim 9: end if 10: else if state = interim then 11: if filtered threshold and counter < maxcount then 12: counter++ 13: else if filtered threshold and counter > mincount then 14: counter 15: else if filtered threshold and counter = maxcount then 16: iscardetected true 17: packet(car=yes) 18: state car 19: else if filtered < threshold and counter = mincount then 20: if iscardetected = true then 21: packet(car=no) 22: state no_car 23: end if 24: end if 25: else if state = car then 26: if filtered < threshold then 27: counter 0 28: state interim 29: end if 30: end if 31: end loop unit. This unit aggregates data from different sensors to determine the speed of each vehicle, and to track it until it leaves the intersection. Determining the velocity of the vehicles can be accomplished by using two sensors, separated by a known distance. Similarly, acceleration can be derived from two consequent velocity estimates. The difficulty in such a system is that, when a car is detected, the corresponding event has to be accurately timestamped. In a network which is not time synchronized, networkrelated delays can be introduced because of retransmissions, buffering or different length of multi-hop paths. This causes a time stamping accuracy in the order of several seconds, which is not accurate enough to estimate acceleration, velocity and position. Standards such as IEEE e are in preparation to replace IEEE [6], and operate in a synchronized manner, allowing for synchronization in the order of tens of micro-seconds. We are currently working on integrating the time-stamping capability of this standard into a smart intersection. VIII. CONCLUSION This paper contributes towards making left turns at intersections safer. We believe that Wireless Sensor Networks can play an important role in achieving that role, as they can be used to detect the presence of vehicles, and to transmit data reliably to a smart intersection management unit. This unit can then trigger the appropriate warnings to the drivers. In this paper, we have designed a sensor board using a 3-axis magnetometer to detect the presence of vehicles. We have developed a threshold detection algorithm that takes the amplitude of the magnetometer as an input, and decides whetheravehicleispresent.wehaveshowntheefficiency of this algorithm by running it off-board on a set of collected magnetometer traces, and have implemented it on-board on a typical micro-controller. REFERENCES [1] Jim Misenser, Cooperative Intersection Collision Avoidance System (CICAS): Signalized Left Turn Assist and Traffic Signal Adaptation, California PATH Research Report, UCB-ITS-PRR [2] Jim Misener, California Intersection Decision Support: A Systems Approach to Achieve Nationally Interoperable Solutions II, California PATH Research Report, UCB-ITS-PRR [3] Ankur M. Mehta and Kristofer S. J. Pister, WARPWING: A complete open source control platform for miniature robots, 2010 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2010), Taipei, Taiwan, Oct [4] HMC5843, Honeywell 3-Axis Digital Compass IC, [5] AT86RF231, MCU Wireless, Atmel Corp., [6] IEEE WPAN Task Group 4e (TG4e), [7] Sacramemto Street at University Avenue, Berkeley, CA, retrieved on 09/27/2010 from website [8] MATLAB version R2009a. Natick, Massachusetts: The MathWorks Inc., [9] Sio-Song Jeng, Patrice Briand, Infrastructure and Vehicle Communication for Speed Limitation Based on Magnetic Markers, Intelligent Vehicles Symposium 2006, June 13-15, 2006, Tokyo, Japan [10] Ching-yao Chan, A System Review of Magnetic Sensing System for Ground Vehicle Control and Guidance, California Partners for Advanced Transit and Highways (PATH), UC Berkeley, [11] Sangwon Lee, Dukhee Yoon, Amitabha Ghosh, Ming Hsieh, Intelligent Parking Lot Application Using Wireless Sensor Networks, International Symposium on Collaborative Technologies and Systems, 2008 [12] Jiagen Ding, Vehicle detection by sensor network nodes, MS thesis, Department of Electrical Engineering and Computer Science, University of California, Berkeley [13] Caruso and Withanawasam, Vehicle Detection and Compass Applications using AMR Magnetic Sensors, Honeywell Report [14] J, F. Scarzel lo, D. S. Lenko, R. E. Brown, A. D. K rall, SPVD: A Magnetic Vehicle Detection System Using a Low Power Magnetometer, Page IEEE Transactions on Magnetics, VOL. MAG-14, NO. 5, September 1978 [15] J Scarzello, G Usher Jr - A Low Power Magnetometer For Vehicle Detection, IEEE Transactions on Magnetics, 1977 [16] Wassim G. Najm, Jonathan A. Koopmann, David L. Smith, Analysis Of Crossing Path Crash Countermeasure Systems, United States Department of Transportation, Volpe National Transportation Systems Center, National Highway Traffic Safety Administration [17] RJ Koerner, Method and apparatus for vehicle detection - US Patent 3,249,915, 1966