TARGET DETECTION AND TRACKING IN A WIRELESS SENSOR NETWORK

1
TARGET DETECTION AND TRACKING IN A WIRELESS
SENSOR NETWORK
Clement Kam, William Hodgkiss, Dept. of
Electrical and Computer Engineering, University
of California, San Diego
Wireless Sensor Network Target Tracking The
wireless sensor network version of the tracking
simulation reports data only when the target
enters into a sensors sensing range. The data
reported is the assumed location of the sensor
which is randomly perturbed from the true
location to model uncertainty in the location of
the sensors.
Abstract Target detection and tracking in a
wireless sensor network is studied. A Kalman
filtering approach is applied both to mitigate
false alarm as well as to smooth noisy target
position estimates. Simulations illustrating the
tracking results are presented. Introduction The
emergence of wireless sensor network technology
has led to applications in many areas, including
industrial, military, and health care fields.
Some of the major issues with setting up and
running wireless sensor networks in these areas
are power conservation, distributed processing,
and goal-oriented, on-demand processing. In the
detection and tracking system presented in this
work, a first-step approach to a tracking system
is presented. A Kalman filter algorithm is used
to track a target passing through a field of
sensors, and the system can predict what area the
target is expected to lie in with high
probability. This is used to ensure that the
detection made by a sensor fits with the targets
motion profile, thus minimizing false detections.
This approach is taken from the paper by McErlean
and Narayanan 1.
Generic Two Dimensional Tracking Problem A two
dimensional tracking scenario using Kalman
filtering was simulated. Here, observations are
available at every time step.
Sparsely Sampled Two Dimensional Tracking
Problem To study the effect of sparsely sampled
data, the Kalman filter tracker was applied to a
case where a section of data is missing.
Figure 5. The sensors have a sensing range of 4
and are indicated by the dotted lines. Their true
locations are indicated by the crosses, and the
squares are the data reported. The assumed
location is perturbed from the true location with
a standard deviation of sk2. The red circles
are the 3s prediction circles, and in this case,
have not failed to predict the region where the
target is located at the moment of detection. As
a mitigator of false detection, the Kalman filter
prediction has confirmed that all of the events
are caused by the target. The small circles are
the Kalman filtered data. Their performance is
difficult to analyze in a quantitative fashion
for such a small number of samples.
Linear Model for Target Motion This is the model
for target motion that will be used for all of
the simulations.
Figure 1. The above shows a tracking simulation
for a target starting at (5,5) plus noise. In
this system, data is available at every time
step.The red circles show where the Kalman filter
predicts the target should be within a 3s radius.
The circles shrink in size asymptotically as time
progresses, since there is greater confidence in
where the target should be.
Figure 3. The above shows the tracking simulation
when data is missing for 10 samples. As a result,
the prediction for where the target is after 10
samples is less certain, hence the large circle.
Although the target cannot be tracked for those
10 samples, the prediction is still able to
enclose the targets location, which is indicated
by the square.
where
w(n) is zero-mean Gaussian white noise with
covariance matrix
Qk
T is the time step, which in this case was
chosen to be 0.1 seconds, and q is some constant.

Basic Kalman Filter Equations
Figure 4. The variance of the prediction
decreases when data is available but grows
sharply in the absence of data. Less information
about the target location leads to less certainty
in the predicted location.
Figure 6. The prediction variance (size of the
red circles in Figure 5) increases with the
number of samples that occur between target
detections (i.e., no data). The connected line
shows the prediction variance at the times of
detection, and the crosses show the prediction
variance in between detections. It can be seen
that following a detection (new data), there is a
sharp drop in the variance, because there is
greater certainty in the location. As time passes
and no new data is available, there is a steep
increase in prediction variance, as the
prediction grows more uncertain with no
observations.
Given a linear model for target motion, the
prediction step in the Kalman filter can be
exploited to predict, with high probability, an
area where the target is expected to be in the
future. The Kalman equations are given as below
Figure 2. The variance shown here is the
predicted variance Pk1- in the x- (or
equivalently, the y-) coordinate. The variance
corresponds to the red circles in Figure 1.
Corrector Equations
Future Work For the wireless sensor network
scenario shown, only a few samples have been
processed in a single simulation. It would be
desirable to run the same simulation for a longer
time in a larger sensor field to obtain more
samples, and be able to study the tracking
ability of the Kalman filter more closely.
Another modification would be to have smarter
sensors that can detect where inside the sensing
range the target is for comparison with the
results shown here. Another issue that concerns
wireless sensor networks is power management,
since the sensing motes tend to have a limited
battery life. The Kalman filter prediction is
used here to prevent false alarms, but it can
also be used to select a subset of the sensors to
be turned on, and for the others to be asleep.
Predictor Equations
Reference 1 D. McErlean, S. Narayanan,
Distributed Detection and Tracking in Sensor
Networks, Conference Record of the Thirty-Sixth
Asilomar Conference on Signals, Systems and
Computers, 2002. Volume 2, 3-6, pp. 1174 - 1178
vol.2, Nov. 2002.