A Vehicle Detection Algorithm Based on Wireless Magnetic Sensor Networks

Transcription

1 2013 8th International Conference on Communications and Networking in China (CHINACOM) A Vehicle Detection Algorithm Based on Wireless Magnetic Sensor Networks Xiangke Guan 1, 2, 3, Zusheng Zhang 1, 3, 4, Jingquan Zhou 2, Fengqi Yu 1 Department of Integrated Electronics, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China 2 College of Electronic Science and Engineering, Nanjing University of Posts and Telecommunications, China 3 Shenzhen Key Laboratory for RF Integrated Circuits, China 4 The Chinese University of Hong Kong, Hong Kong, China xkguan@hotmail.com; zs.zhang@siat.ac.cn; fq.yu@siat.ac.cn; zhoujq@njupt.edu.cn 1, 3, 4 Abstract At present, the research of management system based on Wireless Sensor Networks (WSN) and Anisotropic Magneto Resistive (AMR) sensors have made great progress. However, due to the diversity of the vehicle magnetic signal and the interference caused by the adjacent spaces, vehicle detection technique using wireless magnetic sensor networks is still immature. For accurately detecting a vehicle in a lot, we propose a detection algorithm named Relative Extremum Algorithm (REA). On the lot at Shenzhen Institutes of Advanced Technology (SIAT), 82 sensor devices are deployed to evaluate the performance of REA. By running the system for more than half a year, we observe that the vehicle detection accuracy of the REA is above 98.8%. Keywords- wireless magnetic sensor networks; vehicle detection; Min-max algorithm; threshold-based algorithm; management system I. INTRODUCTION Wireless Sensor Networks (WSN) [1] is an emerging technology and has great potential to be employed in critical situations. The key advantage of WSN is that the network can be deployed easily and can operate unattended, without the need for any pre-existing infrastructure and with little maintenance [2]. Recently, vehicle detection based on magnetic sensor is a hot research topic [3-4]. The magnetometers available today can sense magnetic fields with the geomagnetic field below 1 gauss. They can be used for detecting the vehicles made of ferrous objects that disturb the earth s magnetic field [5]. In this work, we design and implement a management system based on wireless magnetic sensor networks. Each sensor node is equipped with a magnetic sensor. It determines whether a space is occupied or not by detecting the presence of a vehicle based on a change in the environmental magnetic field. Up to now, many algorithms have been proposed for vehicle detection. One of them using a threshold of magnetic field is to measure only the change in z-axis of the geomagnetic field caused by vehicle [6-7]. But a single threshold is not valid to all situations and often leads to a high false alarm rate. In order to solve this problem, an adaptive threshold algorithm is presented [8-9]. However, threshold-based vehicle detection algorithms have congenital defects. Firstly, there is no existence of a reasonable threshold to distinguish between the signal of vehicle and interference, especially in low Signal- Noise-Ratio (SNR). Furthermore, different places have different geomagnetic fields, so it is difficult to have a threshold to fit all situations. Our experiments also show that the aforementioned algorithms have some accuracy issues when they are applied to the real situations. Their vehicle detection success rates are usually less than 95%. Another detection method is to study the characteristics of the wave pattern. J. Ding proposed min-max detection algorithm [10]. Modified min-max algorithm for magnetometers was also published [11]. These algorithms analyze the process of signal changes. However they still rely on single threshold and have low vehicle detection success rate. Furthermore, Li Cui et al. [4] [12] proposed a Similarity Based Vehicle Detection (SBVD) algorithm to detect vehicles in low SNR conditions by calculating the similarity between on-road signal and a reference signal. A data fusion algorithm based on fuzzy logic theory has also been proposed to monitor a space using magnetic sensor [13]. These two algorithms both have high computational complexity and also do not meet the accuracy requirement. In order to achieve a higher detection success rate and lower computational complexity, based on analysis of a large amount of experimental data, we propose a vehicle detection algorithm named Relative Extremum Algorithm (REA) which is based on the min-max detection algorithm. Through multiple threshold setting, it can eliminate the limitations of the single threshold and different decision methods are used for different magnetic signal strength of vehicle. On the space at Shenzhen Institutes of Advanced Technology (SIAT), 82 sensor nodes are deployed to evaluate the performance of REA. By running the system for more than half a year, we observe the vehicle detection accuracy of REA is better than 98.8%. The remainder of this paper is organized as follows: Section 2 describes the characteristics of the magnetic signal. Section 3 proposes the relative extremum algorithm. Section 4 demonstrates our simulation and experiment results. Finally, section 5 makes a brief conclusion. This work was partially supported by the funds of Shenzhen government, with grant numbers: JC A, JC A, NST , CXB A, CXY A, and CXZZ IEEE

2 II. CHARACTERISTICS OF THE SIGNAL The feature analysis and recognition method of geomagnetic signal is the basis of vehicle detection algorithms. However, very few publications have done a detailed analysis of the characteristics of the geomagnetic signal. Based on the original magnetic signal collected from the experiments, the geomagnetic signal is studied and summarized in the time domain. A. The Deployment of Sensor Nodes In our experiments, each sensor node is placed in the middle of a space. We use Honeywell 3-axis magnetic sensor [5] for vehicle detection. Figure 1 describes the deployment of sensor nodes: the Z-axis is vertical, the Y-axis is parallel to the direction of vehicle entering, and the X-axis is pointing to adjacent space. There are typically two layouts, vertical and parallel. The vertical refers to the space perpendicular to the driveway, as shown in Figure 1(a). The parallel refers to the space parallel to the driveway, as shown in Figure 1(b). In our experiments, we found that some of the vehicles greatly disturb the geomagnetic field. They often bring interference to the sensors at the adjacent spaces. For simplicity, we define two kinds of signals: 1) The signal which is perceived by the sensor node deployed on the vehicle parked space. 2) The interference signal which is perceived by the sensors at the neighbor spaces. The signal and the interference signal of a vehicle are shown in Figure 4. Figure 2. The magnetic signature of a vehicle on a vertical space. (a) The vertical. Figure 3. The sensor signal for parallel for a vehicle entering a space with its head forward. (b) The parallel. Figure 1. Two typically layouts and the placement of magnetic sensors. B. Experimental Data Figure 2 shows the magnetic signal for a vehicle on a vertical space. Figure 3 is the magnetic signal for a vehicle on a parallel space with its head forward. From Figure 3, we can see the engine s turning-on and turning-off cause disturbance to the magnetic signal. (a) Parking signal. 670

3 III. RELATIVE EXTREMUM ALGORITHM Figure 6 shows the block diagram of our proposed relative extremum algorithm. The algorithm consists of signal filtering and min-max detection. Signal filtering uses wavelet transform [14-15]. (b) Interference signal. Figure 4. The signal and interference signal of a vehicle. C. Signal Characteristic 1) As shown in Figure 2 and 3, magnetic signal is oscillating when a vehicle enters or leaves a sensor. When the vehicle is completely stopped or left, the signal becomes stable. The drift of the magnetic signal between two stable states can be calculated, denoted by V1, as shown in Figure 5. At present, the existing algorithms are mostly based on the signal level difference to decide whether the space is occupied. However the interference from the vehicles on the neighbor spaces can produce a similar drift, especially in the case of low SNR. As shown in Figure 4(b), a threshold for the determination of interference usually causes misjudgment. From Figure 5 we can see, another variation appears when vehicle is leaving the space, denoted by V2. V1 equals V2 approximately. Figure 6. Block diagram of the proposed algorithm. Figure 7 shows the state machine used in the proposed REA. The state machine detects the waveform characteristics in the process, including local maxima, local minima, signal stable value during and after, the frequency of fluctuation. It consists of: State(x) : {Init, Flat, Peak, Valley, Flat_count_up, Flat_count_down, Count0} Input(u) : {1, -1, 0} Output(d) : {free, occupancy} The input for the decision state machine is the sign of the slope of s(k), which is defined as: 1, if s(k) - s(k - 1) > min_delta_u u(k) = -1, if s(k) - s(k (1) - 1) < -min_delta_u 0, otherwise where min_delta_u is a pre-defined positive constant. Figure 5. V1 equals V2 approximately. 2) We also note that the signal change in Z-axis during is usually greater than that in X-axis and Y-axis. While the change of interference signal in Z-axis is less than the other two. This is because the sensor Z-axis is placed perpendicular to the ground and it is a sensitive axis. However, the sensor Z- axis is insensitive to some special vehicles. 3) Generally, the change of interference signal is relatively flat, as shown in Figure 4(b). The interference caused by moving vehicles on the road produces a short fluctuation, and does not cause a stable drift. According to our statistics, most of the vehicles have these two features which are the basis of our proposed algorithm. Figure 7. State machine for REA detection. There is a counter associated with each state in {Peak, Valley, Count0}, to introduce hysteresis in the detection. When the machine jumps from one state to a new state, the counter associated with the new state resets and only counts up when the state loop back to itself. 671

4 Flat: After system initialization, the state machine jumps to Flat state and stays at this state if u(k)=0. The state machine can jump to Flat_count_up state when the current slope is positive (u(k)=1) and jumps to Flat_count_down when the current slope is negative (u(k)=-1) accordingly. Flat_count_up: The state machine stays where it is if the slope is positive (u(k)=1) and it jumps to Count0 if the current slope is equals to 0 (u(k)=0). Moreover, the state machine jumps to Peak state if the current slope is negative (u(k)=-1). Flat_count_down: The state machine stays where it is if the slope is negative (u(k)=-1) and it jumps to Count0 if the current slope is equals to 0 (u(k)=0). Moreover, the state machine jumps to Valley state if the current slope is positive (u(k)=1). Count0: The state machine stays where it is and the counter counts up if u(k)=0 and the counter has the value less than N. The state machine jumps to Flat if u(k)=0 and the counter has the value not less than N. Moreover, the state machine jumps to Flat_count_up if u(k)=1 and jumps to Flat_count_down if u(k)=-1. Peak: The state machine stays at Peak and the counter counts up if the slope is not positive (u(k)=-1) and the counter has the value less than M. The state machine jumps to Flat_count_down if the slope is negative (u(k)=-1) and the counter has the value not less M. The state jumps to Valley state if u(k)=1. Valley: The state machine stays at Valley and the counter counts up if the slope is positive (u(k)=1) and the counter has the value less than M. The state machine jumps to Flat_count_up if the slope is positive and the counter has the value not less M. The state jumps to Peak state if u(k)=-1. There are also five variables associated with the state machine: slope_change_counter, local_min, local_max, flat_previous and flat_current. They are updated as follows, where x is the current state. s(k), if x {Flat, Count0} flat_current = unchanged, otherwise The local_min tracks the local minimums in real time while the local_max tracks the local maximums. The slope_change_counter indicates the number of the magnetic signal fluctuations. While flat_previous - flat_current indicates the drift between two stable status, as shown in Figure 8. Occupancy of space is determined by the flow chart, as shown in Figure 9. δ and α represent the fluctuation of the signal frequency and the relative amplitude, respectively. In order to distinguish the amplitude of the vehicle magnetic signal, the algorithm sets the high and low threshold U_max and U_min. Different decision is used for different magnetic signal strength. Thereby the detection accuracy is improved. Figure 8. Variable update process. (6) slope_change_counter(k - 1) + 1, if x {Peak, Vally} slope_change_counter(k) = 0, if x {Flat} unchanged, otherwise min{s(k), local_min(k - 1)}, if x {Valley} local_min(k) = s(k), if x {Flat} unchanged, otherwise local_max(k) = max{s(k), local_max(k - 1)}, if x {Peak} s(k), if x {Flat} unchanged, otherwise flat_previous = s(k), if x {Flat} unchanged, otherwise (2) (3) (4) (5) Figure 9. The flow chart of the proposed algorithm. 672

5 IV. EXPERIMENT AND DATA PROCESS In order to verify the proposed algorithm, we deployed 82 nodes in the lot at Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences. The system has operated for 6 months. According to the large number of measured data, we found that the two coefficients of the proposed algorithm, δ and α, should be 3 and 0.8, respectively. The U_max and U_min should be 30 and 5, respectively. TABLE I lists the coefficients in REA. TABLE I. PARAMETER IN REA. Parameter Value Parameter Value min_delta_u 3 U_min 5 M/N 10 δ 3 U_max 30 α 0.8 Three-axis magnetic field data are analyzed using fuzzy logic and simulated using Matlab. The detection successful rate is above 98.8%. Figure 10 shows the sensor signal process and decision of space occupancy. The first picture shows the filtered signal s(k). The second and the third show the process of the changes of local maxima and minima. The fourth represents the update of the fluctuation of signal frequency. The last picture shows the results using the proposed algorithm. 1 means the spaces occupied and 0 means not. The experiment results are given in TABLE II-IV. of TABLE I and TABLE V, we can see that the REA is better than the threshold-based algorithm. TABLE III. Location of sensor No car on left & right A car on the left A cat on the right 2 cars on left & right TABLE IV. Parking style Head-in Parking Back-in Parking TABLE V. VEHICLE PARKING DETECTION USING REA: GROUP1. times of detected vehicle Accuracy of the proposed algorithm VEHICLE PARKING DETECTION USING REA: GROUP2. times of detected vehicle Accuracy of the proposed algorithm VEHICLE PARKING DETECTION USING THE THRESHOLD-BASED ALGORITHM. Detected the number of Accuracy of the threshold-based algorithm It is noted that the counter at the states {Count1, Peak, Valley} and parameters M and N introduce hysteresis in the detection, which makes the algorithm more robust to the short burst errors in the hard decision. Figure 10. The sensor signal process and decision of space occupancy. TABLE II. VEHICLE PARKING DETECTION USING REA. times of detected vehicle Accuracy of the proposed algorithm For comparison, some experiments have also been done to evaluate the performance of the threshold-based algorithm. The experimental results are shown in TABLE IV. By comparison V. CONCLUSION In this paper, we have studied an Intelligent Parking System using wireless magnetic sensor network. Based on the analysis of a large number of experimental data, we have proposed the Relative Extremum Algorithm (REA), to solve the limitations of the existing threshold-based vehicle detection algorithm. Our simulations and experiments demonstrate that the proposed algorithm can accurately detect space occupancy. The system is stability and has high detection accuracy. It has been working in the lot at SIAT for 6 months. REFERENCES [1] C. S. Raghavendra, K. M. Sivalingam, and T. Znati. Wireless sensor networks. Kluwer Academic Pub, [2] G. Padmavathi, D. Shanmugapriya, and M. Kalaivani. "A Study on Vehicle Detection and Tracking Using Wireless Sensor Networks," Wireless Sensor Network, vol. 2(2), 2010, pp [3] M. J. Caruso and L. S. Withanawasam. "Vehicle detection and compass applications using AMR magnetic sensors," Sensors Expo Proceedings, vol

6 [4] Z. lei, R. Wang, and L. Cui. Real-time Traffic Monitoring with Magnetic Sensor Networks, Journal of information science and engineering, vol. 27(4), 2011, pp [5] Information on [6] S. Coleri, S. Y. Cheung, and P. Varaiya. "Sensor networks for monitoring traffic," Allerton conference on communication, control and computing, 2004, pp [7] A. Haoui, R. Kavaler, and P. Varaiya. "Wireless magnetic sensors for traffic surveillance," Transportation Research Part C: Emerging Technologies, vol.16(3), 2008, pp [8] J. Ding, S. Y. Cheung, et al. "Signal processing of sensor node data for vehicle detection," Intelligent Transportation Systems, Proceedings. The 7th International IEEE Conference on. IEEE, 2004, pp [9] S. Y. Cheung. Traffic Surveillance by Wireless Sensor Networks. Phd thesis, University of California, Berkeley, [10] J. Ding. Vehicle detection by sensor network nodes. MS thesis, Department of Electrical Engineering and Computer Science, University of California, Berkeley, CA, Fall [11] S. Lee, D. Yoon, and A. Ghosh. "Intelligent lot application using wireless sensor networks," Collaborative Technologies and Systems, CTS International Symposium on. IEEE, 2008, pp [12] Z. Yuhe, H. Xi, et al. "Design and Evaluation of a Wireless Sensor Network for Monitoring Traffic," The 14th World Congress on Intelligent Transportation Systems (WCITS'07) [13] Z. Jian, H. Cao, et al. "Data Fusion for Magnetic Sensor Based on Fuzzy Logic Theory," Intelligent Computation Technology and Automation (ICICTA), 2011 International Conference on, vol. 1, IEEE, 2011, pp [14] I. Daubechies. "The wavelet transform, time-frequency localization and signal analysis," Information Theory, IEEE Transactions on, vol , pp [15] M. Alexandrescu, D. Gibert, et al. "Detection of geomagnetic jerks using wavelet analysis," Journal of geophysical research 100.B7, 1995, pp