Road Traffic Monitoring Using A Wireless Vehicle Sensor Network
Road Traffic Monitoring Using A Wireless Vehicle Sensor Network
Road Traffic Monitoring Using A Wireless Vehicle Sensor Network
Hf'a\'yVehicle
Fig. 3: Vehicle sensors node outputs
The PIR sensor is used together with a Fresnel lens and
interfaced to an amplifying circuitry. PIR sensor was used to
detect the oncoming vehicles from a further range of 2--3m
as compared to the other sensors. This can be used for energy
saving features e.g. to turn on other near-field sensors only
when a vehicle approaches. The disadvantage of this sensor
is its inability to differentiate vehicles from other heated
objects like humans or animals. An example of the sensors'
outputs for "No vehicle", "Light vehicle" and "Heavy
vehicle" can be visualized in Fig. 3.
As seen from Fig. 3, the acoustic waveform is able to
detect oncoming vehicles earlier as opposed to near-field
sensors like the magnetic sensor and accelerometer. The
converse is also true in that it is the last sensor to return to
'rest' after the vehicle leaves the area of interest. Thus this
encapsulating effect of the acoustic sensor (over other
sensors) is used as a detection window for vehicle
monitoring in this project. It is only during this detection
window will the remaining sensors be monitored. The
recorded readings from other sensors will be used as validity
check after the window is terminated. This will allow the
end-node to identify false triggering by checking whether
each sensor is 'disturbed' during the window period. For
example a strong gust of wind can trigger an acoustic
detection window but the PIR sensor can be used to
invalidate the detection. Thus the co-operative use of
multiple sensors for verification can reduced the number of
false detection. Each end-node only makes use of four raw
signals namely PIR, acoustic, magnetic X-axis and
accelerometer Y-axis as opposed to the possible seven. This
is because the waveforms of both magnetic axes are very
similar and thus only one axis (X-axis) is selected for used.
Similarly one out of the three axis of the accelerometer is
used for monitoring. From Fig. 3, accelerometer output X
and Z axes can be observed to produce outputs that are less
significant in magnitude than axis Y. Thus this axis output
was selected to represent the accelerometer readings in the
application. In conclusion, the end-node will be collecting
the following data for transmission to the base node for each
valid detection window.
1. Start and End time of the acoustic signal detected
(thereby period can be calculated).
2. Number of peaks of the PIR sensor within the
window.
3. Minimum and Maximum value of the magnetic
sensor X-axis (thereby peak-to-peak amplitude can
be calculated) within the window.
4. Minimum and Maximum value of the accelerometer
sensor Y-axis (thereby peak-to-peak amplitude can
be calculated) within the window.
A valid detection window with the accompanying captured
data will be forwarded to the base node and subsequently to
the PC. The acquired data from the two sensors nodes will be
fed into a back propagation classifier to determine the type of
vehicle and direction of travel will be predicted.
III. METHODS
The system was setup along Nanyang Avenue at Nanyang
Technological University using 1 laptop and 3 wireless
nodes (2 for end-nodes and 1 for base node). The system
deployment can be seen in Fig. 4. The two end-nodes were
positioned on the curb of the road edge while the
accelerometer of each node was placed on the road side. This
placement of sensor nodes was chosen because it provides
the closest proximity with the vehicles without
compromising the safety of the nodes and vehicles.
The rational for this specific setup is to effectively
monitor both lanes of the road without loss of precision for
either lane. Initially both the end-nodes were placed on the
same side of the road to perform two-lane monitoring due to
ease of deployment. However test results showed that the
sensors reading for the further lane were relatively smaller in
magnitude. This is especially true for magnetic sensor as
readings drop significantly when the detection range
increases. This can be verified by the Honeywell magnetic
sensor application note which states that the sensor output
may drop to a tenth in magnitude if the detection range is
increased by one meter. This presents a problem to the
classifier as sensor readings of the farther lane were not very
representative and more classification error were resultant.
The adopted setup is able to monitor each lane of the road
and thus resolved the above mentioned problem. With this
approach, the end-node that is near the lane of travel will
pick up a stronger signal and thus will have more influencing
power during the classification process.
Fig. 4: Vehicle monitoring system deployment
IV. EXPERIMENTAL RESULTS AND DISCUSSION
RESULTS WITH TWO END-NODES SYSTEM
The system is deployed and tested during the day. Vehicles
traveling on the two-lane road were observed mostly to be of
medium type and of approximately 1.5 vehicles per minute
on both lanes of the road. The test was carried out to gather
50 vehicle readings which consist of heavy, medium and
light vehicles. The system output for the field test is
summarized as shown in Table 1.
The system is able to classify the vehicle type and
direction of travel with reasonably satisfactory accuracy as
shown in Table 1. It should be known that the accuracy of
the classification are tabulated based on the results of each
measured field e.g. Light vehicle, Right to Left Direction.
The neural network is able to achieve 100% classification
accuracy for medium vehicles as opposed to the other two
classes of vehicles. This may be because large portions of the
accumulated training data are of medium vehicles class as
the observed frequency of these vehicles are higher. With
more training data of medium vehicle class, the classifier is
able to build a more accurate classifier for such vehicle type.
Both the light and heavy vehicles are correctly classified to
an accuracy of 70-80% which was observed to vary from
their medium class counterpart.
Many heavy class vehicles are observed to be classified
into the medium class during the field test. This could be
because the sensor readings for both classes are not wide
varying enough. Most of the heavy vehicle training data and
test samples are of middle-sized lorries which are only
slightly bigger in size than conventional Multi-Personnel and
Sports-Utility Vehicles. However if double-decker buses or
other heavy-load vehicles such as dirt truck were included in
the training data and test, the classifier will be expected to
achieve more accurate results. Overall, the vehicle type
classifier is able to achieve 90% accuracy in detennining the
3 classes of vehicles. This high confidence level is partly
attributed by the medium type classification success.
Parameter
Actual S stem Outl ut
Accuracy
Events Light Med Heavy
Light 12 9 3 0 75.0%
Medium 26 0 26 0 100%
Heavy 12 0 2 10 83.3 %
9 31 10
Parameter
Actual System Output
Accuracy
Events Right to Left Left to Right
Direction
24 19 5 79.2tlo
rRight-Leftl
Direction
26 2 24 92.3tlo
[Left-Right]
21 29
Both parameters
50 39
correct 78%
Overall system accuracy (for both parameters)
System accuracy for classifying vehicles into their classes 90%
System accuracy for determining the direction of travel 86%
Overall system accuracy (for each parameter 88%
Table 1: Results of vehicle monitoring system
Directional analysis of the vehicle monitoring system
was also observed to perfonn satisfactorily. The classifier is
able to correctly identify vehicles traveling from right-to-Ieft
and vice-versa with 79% and 92% accuracy respectively.
Logically, the classifier will be able to co-relate the direction
of travel by identifying the sensor node with the higher
significant readings as the incoming direction. The classifier
should be able to utilize this trend to easily detennine the
direction of travel. However this method can only work with
the assumption that both sensor nodes are placed at a specific
distance apart. The system is able to achieve an overall
direction analysis of 86% with the current setup. Out of 50
vehicles that are detected in the field test setup, the system
was observed to correctly classify 39 vehicles into their
respective vehicle class and direction of travel. This 78%
system accuracy indicates the feasibility of the proposed
system for vehicle type and direction analysis. However
more work needs to be done to increase the classification
accuracy.
From the results of the experiment, the type of vehicle
and direction was detennined to be within 90% and 86%
accuracy respectively. The sensors used in the monitoring
node showed correlation to the type of vehicle detected.
However the differences between medium and light vehicles
are less obvious and this resulted in a lower classification
percentage for light vehicles. Nevertheless, the sensors were
able to identify the vehicle and reduce false triggering
through co-operative sensing.
Multiple sensor node system was also identified to
increases the detection accuracy as compared to single node
monitoring. Test results indicate that the use of two end-
nodes provides better classification results in both vehicle
type and directional analysis. The multiple sensor node
system saw a 22% increase in classification accuracy for
both parameters. However this assumption only holds for
single and 2 end-nodes system as verification of 3 and more
end-nodes system are not carried out.
A separate experiment to verify the positive correlation
of increase sensors with classification accuracy was also
carried out. The classification results show an improvement
of approximately 25% in correctly identifying both the type
and direction, as compared to the reduced sensor system.
Through this experiment, we are able to prove that the
combination of magnetometer and accelerometer perfonn
much better than using either of the sensors alone
v. SUMMARY
In this paper, we have devised a multiple nodes vehicle
sensing networked system using a diverse sensor array for
cooperative sensing. This system allows the software-based
classifier to achieve higher prediction accuracy while
reducing false triggering. The use of multiple sensors and
nodes has been demonstrated to produce results that are more
accurate through numerous experimental studies.
Nevertheless more tests need to be done to detennine the
optimal number of sensors (inclusion/removal of sensors)
and nodes to be used. Future work will focus on the effort to
detennine the positioning of each sensor node and how
different configurations can affect the overall system
accuracy. This will assist our understanding for potential
aerial based deployment where the nodes placements are
unstructured. The other interesting area is the use of sensor
array for knowledge discovery. Currently the data from the
four sensors are used only for vehicle-type classification and
travel direction analysis. The acquired sensors signals for
different vehicles can be further processed to extract other
features to provide more detailed infonnation about the
detected vehicle (e.g. speed of traveling). The system will
also be expanded for deployment on aT-junction or
crossroad junction to effectively monitor the traffic
movement. Such a system can eventually be used for
unmanned surveillance that can be deployed aerially.
REFERENCES
[1] T. Haenselmann, "Sensornetworks", http://www.informatik.uni-
mannheim.de/-haensel/sn book!, 5 April 2006.
[2] Q.S. Shan, Y. Liu, G. Prossec, D. Brown, "Wireless Intelligent Sensor
Networks for Refrigerated Vehicle", IEEE 6th CAS Symp on
Emerging Technologies: Mobile and Wireless Comm, Shanghai, China,
May 31-June 2, 2004.
[3] K. Romer, F. Mattern, "The Design Space of Wireless Sensor
Networks", IEEE Wireless Communications Vol: 11 Issue: 6, Dec,
2004.
[4] T.H. Arampatzis, J. Lygeros, S. Manesis, "A Survey of Applications of
Wireless Sensors and Wireless Sensor Networks, Intelligent Control",
Proceedings of the 2005 IEEE International Symposium on,
Mediterranean Conference on Control and Automation, 2005.
[5] Y. Wang, G.R. Zhou, T. Li, "Design of a Wireless Sensor Network for
Detecting Occupancy of Vehicle Berth in Car Park", Seventh
International Conference on Parallel and Distributed Computing,
Applications and Technologies (PDCAT'06), pp. 115-118,2006.
[6] Honeywell, "Vehicle Detection Using AMR Sensors", Application
Note-AN218, 2005.
[7] The Ohio State University, "A Line in the Sand",
http://www.cse.ohiostate.edu/siefast/nest/nest webpage/ALinelnTheSa
nd.html.
[8] M.L. Moran, D.G. Albert, "Source Location and Tracking Capability
ofa Small Seismic Array", U.S Army Corps of Engineers, 1996.
[9] H.D. Wu, M. Siegel, P. Khosla, "Vehicle Sound Signature Recognition
by Frequency Vector Principal Component Analysis", IEEE
Transactions on Instrumentation and Measurement Vol 48 No.5,
October 1999.