CBNP: A Vehicle Monitoring Method for Wireless
Lei Zhang, Rui Wang, Li Cui
Institute of Computing Technology of the Chinese Academy of Sciences
No.6, South Kexueyuan Road,HaiDian, Beijing, China 100190
sink to accomplish monitoring task, however, this will result Abstract—Wireless sensor networks have been proven to be in overconsumption of battery power and overload of radio attractive in intelligent transportation system. To apply wireless communication. So the vehicle monitoring mechanism, which sensor network in vehicle monitoring, two important problems,
vehicle detection and vehicle response feature extraction, should can produce accurate node-level information and effectively be explored. To address these problems, challenges like strong reduce data load even finish detection task, should be noise disturbance, severely limited node resource, should be implemented in sensor node. overcome. In this paper, we propose a CBNP algorithm to detect In magnetic vehicle monitoring systems, considering low vehicle and accurately extract vehicle response feature. We use power consumption and communication overhead, individual cross-correlation to reduce the uncertainty caused by noise or sensor node should adopt an accurate vehicle monitoring disturbance. On the basis of cross-correlation output, we propose mechanism. Based on vehicle monitoring mechanism, and prove a theorem which depicts the relation between vehicle unconcerned data will be dropped, accurate vehicle response feature point and cross-correlation feature point. To information should be extracted and sent to cluster header or support this theorem, we build a theoretical response model for
cross-correlation calculation. Due to the ubiquitous background sink so that the high-level information processing can noise, the direct identification of response feature point is not accomplish system level objectives, such as classification and reliably accurate, especially when Signal-to-Noise Ratio(SNR) is tracking. low. But the identification of cross-correlation point is easy and is To realize an accurate vehicle monitoring mechanism, a relatively insensitive to noise. Simulation also indicates that fundamental problem, feature extraction, should be taken into CBNP outperforms traditional methods on detection accuracy consideration. As in magnetic sensor monitoring systems, and feature extraction accuracy and it works well under lower vehicle feature can be described by several points[6, 8–10], of SNR. response. The start point and end point identify the time
I. INTRODUCTION instant of vehicle entrance and leaving. The peak point
identifies the time instant of strongest response. These basic Magnetic information has proven to be valuable in many information can be used in information fusion to gain high-scientific fields, such as bioelectronics, geology and level vehicle information, including classification, tracking physics. To get a more comprehensive observation from and prediction. Vehicle detection trigger feature extraction to the objective, magnetic information is often continuously calculate the vehicle response feature. To accurately extract monitored by unattended sensors. These continuous vehicle feature, novel methods should also be adopted to monitoring system presents several challenges. Among these reduce uncertainty caused by noise or method itself. The challenges, effective monitoring mechanism and extracted information will be transmitted to upper level environmental noise can not be ignored. So solutions should information fusion node. be explored to deal with these challenges. In many monitoring At present, there is still challenge in feature extraction of applications, distributed wireless sensor networks(WSN) vehicle monitoring mechanism. Due to resource limitation and turn out to be ideal solutions, because of their deployment low capacity, it is difficult to provide accurate node-level convenience, low cost, relatively long self-powered lifetime feature information for upper level fusion computation. When and unattended nature. SNR is low, feature accuracy is not reliable, the upper level Among monitoring related applications, vehicle fusion result is not meaningful. monitoring systems[5, 6] can improve urban transportation In our previous work, we have solve the reliably efficiency, therefore much more attention has been paid to this vehicle detection problem when SNR is relatively low. Based field. When WSN is applied to vehicle monitoring, magnetic on this work, in this paper, we propose an accurate vehicle information is often sensed for vehicle surveillance. In feature extraction mechanism, namely CBNP, to accurately magnetic sensor network for vehicle monitoring, the sensor extract vehicle feature. nodes are supposed to collect environmental measurements Our contributions are as follows: We propose a cross-periodically and transmit measurements to their neighbors or correlation based vehicle detection method. And we build a sinks. To realize real-time vehicle monitoring applications, the reasonable response model for vehicle response used in cross-sampling frequency can be as much as more than 100Hz. correlation calculation. To accurately extract vehicle feature, Consequently, vast sums of data will be produced. we propose and prove a theorem that depicts the relation Theoretically, all the sensory data should be transmitted to
between response feature point and its cross-correlation this section we propose a cross-correlation based vehicle feature point. Finally, we propose a CBNP algorithm to detection algorithm, then we build vehicle response model, accomplish feature extraction. CBNP can accurately extract which is critical in cross-correlation calculation.vehicle feature even when SNR is relatively low. A. Detecting vehicle by cross-correlation
II. RELATED WORK Due to the ubiquitous background noise, when SNR is
relatively low, the MAT method and LPFT method can not Vehicle detection by magnetic sensor network was
effectively reduce the uncertainty, so the feature extraction proposed in . In this paper, a simple threshold-based
accuracy is not reliable. Fig.1 shows a strong response and a vehicle detection algorithm was proposed. Vehicle will be weak response sampled in the direction of X-axis when a announced if sensor reading exceed a threshold. In , vehicle passes an sensor node. Ideally, vehicle detection task vehicle detection is also accomplished by a simple threshold, can be accomplished by processing the raw magnetic and the threshold can be dynamically adjusted by a CFAR data(Fig.1a). However, relative strong noise or low SNR detector. Based on the detection result, the sensor node
(Fig.1b) always results in high false detection rate and low extracts three vehicle response feature: duration that the
detection rate. detector exceeds the threshold, maximum sensor reading and
its time instant. Researchers in  proposed similar CPA
algorithm to accomplish detection task and extract vehicle
feature points from the raw signal with the help of an
adaptable threshold. However, these work had not paid
enough attention to noise, so slight SNR decline will lead to
low system performance.
In , the sampled sensory data is first processed by a
moving average statistics(MAS) to reduce noise, then a
threshold is used to determine whether an vehicle appears.
After that, vehicle feature can be calculated. However, this
method also suffers from low S(NR.
In , the noise is reduced by a digital filter (low pass
filter), the detection task is then performed. In , the raw signal is first input into a limiter and then a low pass filter to reduce noise, after that , decimator, moving statistics, CFAR
and energy estimator will be imposed to the signal and finally
some features of vehicle response signal, such as duration, Fig. 1. Different response caused by vehicle (a.strong b.weak ) energy, will be calculated. However, the first two steps, limiter and low pass filter, can not effectively reduce noise in We solve the vehicle detection problem from a different all cases. Furthermore, the low pass filter will result in phase perspective. We suppose that there is a vehicle response deviation, so the following steps may build on an inaccurate model which is similar to most vehicle response. We input. transform the vehicle detection task to calculate the similarity All the aforementioned methods are mainly based on three between real vehicle response and vehicle response model. basic mechanisms: 1. Direct thresholding. A threshold is used Here we choose cross-correlation as the similarity to detect vehicle. 2. Moving Average and Thresholding measurement. Suppose that the vehicle response model is h(t). (MAT). MAT use an average of neighboring data to replace The real vehicle response is H(t) and H(t) = s(t) + n(t), where original data so that noise disturbance can be reduced. Then s(t) is the ideal vehicle response and n(t) is noise. As the ideal threshold is used to detect vehicle. 3. Low Pass Filter vehicle response is high-correlated with vehicle response Thresholding (LPFT). LPFT use a low pass filter to process model, while the noise is irrelevant with the response model, raw data and then threshold is used to accomplish detection so after correlation calculation, corr(h(t), s(t)) will reach peak tasks. Direct thresholding has been proven to be not fit for real while corr(h(t), n(t)) remain low(as indicated in (1)). application environment. To some extent, MAT and LPFT can (1) corrhtHtcorrhtstcorrhtnt((),())((),())((),())，；be implemented in real application. On the basis of vehicle
detection algorithm, feature extraction algorithm locates B. Vehicle response model feature point. The feature extraction accuracy directly decides
In (1), the first and critical step of correlation calculation is high-level information fusion quality, however, noise always
the choice of vehicle response model. In this part we will results in inaccurate feature calculation.
discuss the vehicle response modeling and then build a vehicle III. CROSS-CORRELATION BASED VEHICLE DETECTION response model. ALGORITHM When a large ferrous object pass an observation point In CBNP mechanism, accurate feature extraction should nearby, the earth magnetic intensity will be changed. Fig.2 be built on effective vehicle detection, because if vehicle is shows a scene that a vehicle pass an observation point(sensor). not detected, feature extraction can not be triggered at all. In
We suppose the time instant is T = 0. We also suppose that md3?vehicle proceeds at the horizontal direction, its speed is v (and ||t？3?vwe assume speed to be invariable when vehicle is in the 2222ht()， (5) ?(()dvt；sensor’s probing range), the vertical distance between vehicle ?and sensor is d and the horizontal distance between vehicle 0others?and sensor is l(We suppose l is large enough). According to It should be pointed out that in this model, d and v should , vehicle-caused magnetic intensity along the horizontal be determined according to different applications. For direction around sensor can be expressed by the following example, if the CBNP algorithm is supposed to accomplish equation: feature extraction in an express way, v should be close the mmost vehicles run on that road. In another case, if the (2) h，3application is deployed on an urban road where vehicle can urnot go so fast as in express way, the v parameter should be correspondingly adjusted. d parameter should be adjusted according to the deployment details, as the sensor node is
commonly deployed alongside a road, the determination of d
should take the lane width into consideration. When a real environment is considered, the noise can not
be ignored. Here we suppose noise to be a gaussian white ；
2noise with zero mean and variance. So the modeling of ；
2noise is to identify the parameter. We will determine the ；
value of in simulation section. So real vehicle response ；
can be expressed by: Fig. 2. Vehicle and sensor. m (6) Htt()()，；； 3In (2), m is earth pole strength, which is a constant, r is the 222(())dvt；(distance between vehicle and sensor, μ is a media constant. At
time instant T = t, distance between vehicle and sensor can be IV. CROSS-CORRELATION BASED VEHICLE DETECTION expressed by: ALGORITHM 22As in sensor node, the cross-correlation calculation should (3) rtdlvt()()，；？
be implemented in a discrete form, as expressed by (7). In (7), According to (2) and (3), the magnetic intensity at time f is a sampled data series, g is data series of vehicle response instant t can be expressed by the following equation: model, g lasts a duration of M. mM？1 (4) ht()，3ftmMgm()()；？~222(((()))：；？dlvtm，0ftg()*， (7) MM？？1122ftmMgm()()；？~~mm，，00
Based on the cross-correlation properties, we have the following theorem:
Theorem1: Given the response model, is the time instant tp
of vehicle response peak,is the time instant of its correlation ts
peak, Then, is constant, and ttM？，/2, where M tt？spsp
is the length of response model.
Proof: In appendix.
Fig.4 shows the relation depicts by theorem1. Fig. 3. Ideal vehicle response. Now we propose CBNP feature extraction algorithm. The ideal vehicle response is showed in fig.3. From ?g.3 When the cross-correlation peak is located, the corresponding lraw data peak can easily be located. This is the basic idea of we can see that the vehicle response is symmetric about . t，vCBNP algorithm. As in sampled real data, noise is included. As the vehicle response is very weak in most time duration, so However, the cross-correlation result is rarely disturbed by ld3noise. Fig.5 is a cross-correlation result of real vehicle here we use a threshold (as displayed in fig.3) to ||t？？vvresponse. From this figure we may find that, due to relatively build vehicle response model. So the vehicle response model strong noise, with traditional methods, it is really difficult to can be expressed by following equation(l is removed):
locate the peak of vehicle response. The cross-correlation to 14, 1 as step) to generate real vehicle response described in
2output is much smoother, so the cross-correlation peak can be of (6). According to the on-road sensory data, we set ；easily located and the vehicle peak can be accurately located. noise to be 11.5 (unit is HMC1000’s resolution). For each run, It should be pointed out that, in the cross-correlation we specify a value for d and v, then 1000 vehicle response calculation and the proof of theorem 1, the vehicle response will be generated and each of them will be processed by three model should be similar to most ideal vehicle response. So we mechanisms (CBNP, MAT, LPFT). Then we will give will discuss how to build a vehicle response model in the next statistical results. part. The MAT algorithm we adopted in our simulation will use the mean of 5 time instants (current, previous two and post
two instant) value to replace current instant value, the mean
replacement operation will be conducted three times. By
this‖smooth‖ operation, the uncertainty caused by noise can
be reduced. The LPFT filter we adopted is a low pass filter, its
pass frequency is 4Hz and its stop frequency is 9Hz. When
CBNP algorithm is referred, as we have mentioned, response
model should be similar to most vehicle response, so here we
choose d = 4 and v = 8.
B. Vehicle detection accuracy To verify the vehicle detection accuracy, we generate vehicle Fig. 4 Time different between response peak and correlation peak
response with different d and v. When a vehicle response is processed by an algorithm, detection failure will be
announced if no vehicle is detected or more than one vehicle
is detected. Fig.6 shows the overall results of simulation.
From this figure we can see that CBNP detection accuracy
mainly outperforms the rest two algorithms.
Fig. 5 Real vehicle response and its correlation peak
V. SIMULATIONS AND VERIFICATIONS
In this section we verify the feature extraction accuracy of CBNP algorithm. According to (6), the vehicle response is Fig. 6 Detection accuracy mainly influenced by v, d and . In following parts we ； discuss the relation between performance and v, d,. As v ；
C. Feature extraction accuracy and d should be meaningful for monitoring applications, we
numerate all the meaningful value to evaluate performance. In this simulation we will verify the accuracy of peak point As to , we demonstrate the extraction accuracy under ；feature extraction. For each vehicle response, three versions of different SNR. We choose two algorithms to compare with, vehicle response feature will be calculated individually by moving average thresholding(MAT) and low pass filter LPFT, MAT, and CBNP. All the three results will be thresholding(LPFT) algorithm. Finally we verify the feature compared with the actual peak point (predefined by (4)). extraction accuracy based on the on-road sampled data. Standard deviation ！will be calculated, and the accuracy is
measured by deviation rate, , where width is the ！/widthA. Simulation settings actual response width, which can be calculated according to According to the monitoring application, there are several (5). factors should be taken into consideration. As HMC100x Fig.7 shows the extraction accuracy statistics in our is often selected as magnetic sensor in monitoring simulation. By this figure we can see that extraction quality of applications[6, 8] and their probing range is within 6 meters. CBNP is better than that of rest two algorithms in almost all The vehicle speed in urban areas is mainly between 20kph to situations. According to Fig.7 we may see that when the d is 50kph (about 6m/s to 14m/s). So here we use different d small, MAT accuracy is similar to CBNP accuracy, but while (range from 1 to 6, 0.5 as step) and different v (range from 6
the d increases, the MAT accuracy will quickly decreases extraction accuracy as the CBNP method, but there are still while the CBNP accuracy remains in a relative high quality. extra false alarms in MAT method. In fig.8, we use black LPFT is not so accurate because that the low pass filter will arrow to tag false alarm of MAT method and red arrow to tag result in a phase difference which means the feature false alarm of LPFT method. It should be pointed out that the information is distorted. It should be noted that according to false alarm rate of LPFT and MAT methods are higher than Theorem1, the lag between cross-correlation peak and real we demonstrated, because here we only focus on the moments vehicle peak is exactly M/2 , which means the deviation rate is when a vehicle appears. When there is no vehicle, according 0. While in our simulation the deviation rate is larger than 0, to our previous work, disturbance, even noise often cause the reason is that the noise will also account for the cross- false alarms in LPTF and MAT methods.
correlation result. Luckily the similarity between noise and D. SNR impact on Performance response model is much smaller than that between model and
As we have mentioned before, ubiquitous noise leads to vehicle response.
the uncertainty in vehicle feature extraction. And low SNR
always leads to detection failure. In this part we discuss the
relation between SNR and performance. We quantify the SNR
to measure extraction accuracy. It should be noted that in our
following discussion, the SNR is calculated according to the
Fig. 7 Feature extraction accuracy
Next we verify the feature extraction accuracy with on-
road sampled data. A TinyOS version implementation of our
CBNP mechanism is implemented in a sensor node, namely
Easi210, which was put on the roadside. We collected the feature extraction reported by the sensor node, at the same Fig.9. Deviation rate and SNR. time, the raw data was also sent back. It should be pointed out Fig.9 shows the relation between SNR and extraction that we collected the raw data only for the verification quality (deviation rate). As we can see from fig.9, when SNR demonstration. In real deployment, the raw data collection is is large enough, the feature extraction quality of MAT and not necessary. As the sampled data was too long in time, to CBNP is nearly the same. However, as the SNR decreases, the make it clear to demonstrate, in fig.8 we indicates only part of MAT quality quickly get worse but CBNP still remain in a the experiment result, but we believe that all the problems are relatively high quality. LPFT quality is not high at all because clearly demonstrated in this figure. that the low pass filter will introduce a phase difference of
VI. CONCLUSION AND FUTURE WORK
In this paper we proposed a vehicle feature extraction
mechanism, namely CBNP. Our innovation is that we build a
theoretical vehicle response model for vehicle feature
extraction. To accurately extract vehicle feature, we propose
and prove a theorem depicting the relation between feature
point and its corresponding point in cross-correlation result.
Based on the theorem, we propose a high-quality vehicle
feature extraction method. Simulation indicates that our Fig. 8. Feature extraction on real sampled data. proposed method outperforms the existing regular methods, In fig.8, asterisk tags the peak point concluded by MAT especially when the SNR is relatively low. In future work, we algorithm (implemented by MATLAB), square tags the peak will focus on inter-node level information fusion, so that the point concluded by the CBNP mechanism (implemented in an node level information can be used to gain meaningful traffic Easi210 sensor node), round tags the peak point concluded by information. LPFT algorithm (implemented by MATLAB). From fig.8 we
VII. APPENDIX can see that, as we have mentioned before, there is a phase
difference in the LPFT extraction, so the accuracy of LPFT
method is not good enough. The MAT method has the similar
Suppose f is a data sequence which contains a vehicle 2011CB302803, Beijing Natural Science Foundation (BNSF)
under Grant No. 4092045, and the National Natural Science response c, c is express as , as we (,,...,,)cccc121NN？Foundation of China under Grant No.61003292. suppose vehicle response to be symmetric, so .To cc，iNi？；1EFERENCES Rsimplify the case, we suppose the rest part of f is 0. Which  G. Baule and R. McFee, ―Theory of magnetic detection of the heart’s means: electrical activity,‖ Journal of Applied Physics, vol. 36, p. 2066, 1965. fcccccc，，(0,0,...,,,...,,,0,0,...)() J. Dearing, P. Bird, R. Dann, and S. Benjamin, ―Secondary ferrimagnetic 1211NNiNi？？；minerals in Welsh soils: a comparison of mineral magnetic detection methods To be general: , Correspondingly: faa，(,,...)and implications for mineral formation,‖ Geophysical journal international, 12vol. 130, no. 3, pp. 727–736, 1997. aaaacac，，，，，，...0,,,... J. Beyer, H. Matz, D. Drung, and T. Schurig, ―Magnetic detection of 121122ppp；；photogenerated currents in semiconductor wafers using superconducting Suppose g is vehicle response model: quantum interference devices,‖ Applied Physics Letters, vol. 74, p. 2863, 1999. gbbbbbb，，(,,...,,)()1211MMiMi？？； I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, ―Wireless Here we suppose both M and N are even. So, the numerator sensor networks: a survey,‖ Computernetworks, vol. 38, no. 4, pp. 393–422, of cross-correlation (7) at time (t) is : 2002.  R. Bishop, ―A survey of intelligent vehicle applications worldwide,‖ in MM/2Proceedings of the IEEE Intelligent Vehicles Symposium, vol. 2000, USA: ftgababab()*()()，，；~~；？？；；？？；？；itMMiitMMitii111IEEE, 2000. ，，ii11 Y. Zhang, X. Huang, and L. Cui, ―Lightweight Signal Processing in According to Cauchy Inequality: Sensor Node for Real-time Traffic Monitoring,‖ ISCIT07, pp. 1407–1412, 2007. 2222 ababaabb；，；；()() M. Kuorilehto, M. Hannikainen, and T. Hamalainen, ―A survey of iMmitiiitMtiiMi；？？；？；；？？；？；11111application distribution in wireless sensor networks,‖ EURASIP Journal on ―=‖ happens if and only if abab，Wireless Communications itMMitii；？？；？；11and Networking, vol. 5, no. 5, pp. 774–788, 2005. M/2 J. Ding, S. Cheung, C. Tan, and P. Varaiya, ―Signal processing of sensor 2222 ，，；；ftgaabb()*()()~node data for vehicle detection,‖ in IEEE 7th International Intelligent ；？？；？；itMtiiMi11，Transportation SystemsConference, Citeseer, 2004. i1 S. Coleri, S. Cheung, and P. Varaiya, ―Sensor networks for monitoring (‖=‖ happens if and only if ), abab，itMMitii；？？；？；11traffic,‖ in Allerton conference on communication, control and computing, pp. 32–40, Citeseer, 2004. Moreover, ;；，iMbb[1,/2],iMi？；1 D. Li, K. Wong, Y. Hu, and A. Sayeed, ―Detection, classification and tracking of targets in distributed sensor networks,‖ IEEE Signal Processing ，‖=‖ in (8) happens if ;；，iMaa[1,/2],itMti；？？；1Magazine, vol. 19, no. 2, pp. 17–29, 2002.  L. Zhang, R. Wang, and L. Cui, ―Real-time Traffic Monitoring with Moreover, ;；，iNcc[1,/2],iNi？；1Magnetic Sensor Networks,‖ Journal of Information Science and Engineering, Accepted to appear. (,,...)acac，，pp；；1122 K. N. R, Improving Reliability of Wireless Sensor Networks for Target Tracking using Wireless Acoustic Sensors. Obviously, when tpMN，；；/2/2 L. Gu, D. Jia, P. Vicaire, T. Yan, L. Luo, A. Tirumala, Q. Cao, T. He, J. Stankovic, T. Abdelzaher, et al., ―Lightweight detection and classification for then, ,which means, ;；，iMaa[1,/2],itMti；？？；1wireless sensor networks in realistic environments,‖ in Proceedings of the 3rd is maximized. ftg()*international conference on Embedded networked sensor systems, p. 217, ACM, 2005.  A. Arora, P. Dutta, S. Bapat, V. Kulathumani, H. Zhang, V. Naik, V. As the vehicle response peak, so the time tpN，；/2Mittal, H. Cao, M. Demirbas, M. Gouda, et al., ―A line in the sand: A wireless sensor network for target detection, classification, and tracking,‖ Computer delay between correlation peak and vehicle response peak is Networks, vol. 46, no. 5, pp. 605–634, 2004. ? M/2 A. Oppenheim, A. Willsky, and S. Hamid, ―Signals and systems,‖ 1997.  B. Cullity and C. Graham, Introduction to magnetic materials. Wiley-IEEE Press, 2008. ACKNOWLEDGMENT  H. Inc, ―1-and 2-Axis Magnetic Sensora HMC1001,‖ tech. rep., This paper is supported by the National Basic Research HMC1002/HMC1021/HMC1022, 900248 Bey. B. Program of China (973 Program) under Grant No.