Target Tracking in Distributed

Sensor Networks

Ahtasham Ashraf
Presentation Outline

• Objective & Approach

• Elements of a Basic MTT System
• Target Dynamics Estimation
Alpha-Beta Filters, Kalman Filter
• Multiple Target Tracking
Gating, Data Association, NN, JPDA, MHT.
• References
Objective & Approach
• The idea is to track multiple
targets moving through a
distributed sensor network 400.0

using multiple modalities.

• The whole Sensor Network is 300.0

divided into regions which can

be dynamically created & may 200.0
road

overlap with each other. Node1-6

Node15
100.0
• Each sensor of the active Node41-42
Node46-56

regions performs the native 0.0

Node58-61

detections for the targets -200.0 -150.0 -100.0 -50.0 0.0 50.0 100.0

emanating different modality -100.0

signals.
-200.0
Approach Continued…..

• The time series data is used to classify the targets into various
predefined categories, which can be used to assist the multiple target
tracking case.
• The detection & classification results are used by the Localizer in the
Manager Nodes to perform target position estimation after region
detection.
• The tracking algorithm is being used to process the Localizer output to
generate tracks & make accurate predictions for the future target
positions.
• New regions are created as the target escapes the current region’s
surveillance area.
• In this way the target moves through the network & regions are
created.
Elements of a Basic MTT System
• The figure shows the functional elements of a simple recursive Multi-Target Tracking system
• It shows a convenient partitioning of the overall system for introduction but there is a
considerable overlap of the functions of these elements. The distinction between individual
elements becomes less apparent when recently developed techniques like MHT are used.
• Assume this recursive processing is being used & tracks have been formed on previous scans.
• Now the Detection & Localization Algs do the necessary Sensor Data Proc & hence produce a
set of Observations.
• First of all, these observations are gated with current tracks so as to simplify the later
association process. Then more refined Data Association Alg are used to determine the final
assignments.
Continued…

• Observations not assigned to existing tracks can initiate new Tentative Tracks. A
tentative track becomes a Confirmed Track when it satisfies some quality & number of
observation tests.
• Similarly Low quality Tracks, as determined by the update history, are deleted.
• Finally the Track Update Alg is used to predict the track in the future time.
• The Uncertainty in Covariances is used to make the gates for next Data Assoc
recursion.
Target Dynamics Estimation
• The objective of this block is to use the current
measurements & give an accurate estimate of the position
of vehicle & also to predict where the target would be
after certain time T.
• The state estimation has its basics in the principles of
Least Squares Estimation. LSE is basically a batch
processing method that uses multiple scans of data to
estimate parameters that are assumed constant over the
data collection interval. However a recursive form of LSE
can be derived & it leads to Kalman Filtering.
• (70’s) Alpha-Beta (80’s) Adaptive Kalman
(90’s) Interactive Multiple Model (IMM) (00’s
) Non-linear filtering?
• Advantage of KF is to provide a convenient manner to
introduce Process noise to model random target motions.
Continued …

• KF gain is chosen automatically based on assumed target maneuver & measurement

noise model.
K (k )  Pp (k  1) H T  R  HPp (k  1) H T 
1

• Provides a convenient measure of estimation accuracy through Covariance Matrix. This

is helpful in data association stage.
• Can compensate partially for effects of miss-detections by increasing the Prediction Cov
Matrix elements to reflect the expected error association with uncertain data. (JPDA)
• Can use interacting multiple KF in parallel to to track maneuvering targets.
Kalman Filter
The Kalman filter can be used to addresses the general problem of trying to estimate the
state of a discrete-time process that is governed by the linear stochastic difference
equation
X (k  1)  A * X (k )  w(k  1)
X (k )  [ x vx y v y ]T
where the State Transition Matrix A is given by:

1 T 0 0 
 
0 1 0 0
A   

0 0 1 T 
It’s a Constant Velocity model & the target
0acceleration is modeled as white noise w(k)
 0 0 1 
The measurement model
 z(k) = H*X(k) + v(k)
The random variables w(k) and v(k) represent the process and measurement noise
respectively.
They are assumed to be independent (of each other), white and with normal probability
distributions
p(w) ~ N(0,Q) p(v) ~ N(0,R)  
 Q = The process noise covariance matrix
R = The measurement noise covariance matrix
The Process noise Covariance Matrices for the constant velocity filter is given by:

T 3 / 3 T 2 / 2 
Q  2
/
 * m
T / 2 T 
 Q / 02 x 2 
Q / 
0 2 x 2 Q 
where σm is the std dev of maneuver & T is the observation interval.
• These can be easily derived from the equations of motion:
S = vit + ½ at2
Vf = Vi + at
• The Measurement Covariance for the constant velocity filter is given by

 x2 0 
R 2
 0  y

• To start the Filter we have to specify the State Vector X(k) .This can be
specified by proper track Initiation procedures e.g. M/N rule.
• So once the track is established every incoming localization result is
used to update the recursive Kalman Filter to generate an estimate of the
current (Filtered) position & the Predicted position after some time T.
• We keep on monitoring when the target is about to escape the current
region & when that time comes a new region is created & the track is
handed over to the next region.
• In this way the Track ID, State vector & Process Covariance matrices
are propagated along the network as the target moves.
• If we don’t get observations for a track till certain number of
observation intervals, we can drop the track.
 Multiple Target Tracking

• In case of multiple targets, in the same region, the detection &

localization algorithms should be capable enough to sense the
presence of multiple targets & hence give proportional number of (x,y)
observations to the tracking algorithm. Also the classifier should also
be able to identify the vehicles based on their signatures in order to
help the tracker in updating their respective tracks.
• This gives rise to the Data Association Problem. Update Which Track
with Which Observation?
• This problem can be solved by the help of Classifier OR certain
Probabilistic techniques.
Data Association
• GATING: This is the process of
specifying a region within which we are
expecting the next observation. This
helps in eliminating the false
observations to a large extent. For those
observations which fall within the gates,
Data Association Algs are used.
• Different types of gates can be used
depending upon the data behavior e.g
Circular, Elliptical & V Gates.
• The size of the gate is calculated at each
recursion from the Prediction 3 y
Covariance Matrix. So gates grow &
shrink in size due to miss-detections.
• 3 x
Those Tracks which share observations
with each other are grouped together &
are called “Clusters”.
• Data Association Algs are run on these
Clusters.
Data Association ( Techniques….. )
-- Nearest Neighborhood Assoc.
It is the simplest method and maintains
the single most likely hypothesis. A
Hypothesis is an assignment from a set
of Observations to a set of Tracks.
{O1, O2, O3}  {T1, T2}
But the NN approach chooses only one
of the many hypothesis & discards the
others without giving any measure of
the correctness of the decision, which
may be wrong sometimes due to the
measurement accuracy.
X(k+1) = X(k) + K(k+1)*[z(k) – H*X(k)]
Not good with multiple observations case
within a gate.
Auction Alg can be used for solving this
problem.
JPDA (Joint Probabilistic Data Association)
It’s a multiple hypothesis data association
technique developed to handle multiple
targets (as opposed to PDA which was for
single targets), in which the probability of
each hypothesis is calculated by
x y y
P( H )  (  g  ) P x (1  P ) 
i ij D D
i 1
d /2
2

g ij  e ij

(2/  ) M / 2 Si
~ ~
d ij  y S y 1

β = False Alarm density = Pfa/Vg Hyp # Trk 1 Trk 2 Prob

x + y = no of tracks in the hypoth 1 0 0 0.011
x=no of tracks that get obs assigned 2 1 0 0.086
y = no of tracks that don’t get obs assigned 3 2 0 0.053
4 3 0 0.019
So no observation is discarded & all of them
5 0 2 0.041
are used to update the track but with some
6 1 2 0.306
probabilistic weights.
7 3 2 0.068
Residue =  pij[z(k) – H*xi(k)] 8 0 3 0.032
X(k+1) = X(k) + K(k+1)*[Residue] 9 1 3 0.239
10 2 3 0.145
 MHT (Multiple Hypothesis Testing)

• JPDA has the disadvantage in case of

very close hypothesis.
• A better approach is to use a Delayed
Decision Logic so that hypothesis are
propagated in anticipation that
subsequent data will resolve the
uncertainty.
• The Track Oriented MHT Approach
deals with the tracks & adds or
deletes(prunes) tracks on each scan.
• One should expect a potential
combinatoric explosion of hypothesis ,
but there are different techniques to
prune them at each scan.
• Its computationally expensive but can
work in complex cases.
References

• Estimation & Tracking: Principles & Techniques. By Yaakov Bar-

Shalom and Xiao-Rong Li
• Design and Analysis of Modern Tracking Systems. By Blackman and
Robert Popoli
• Applied Optimal Estimation. By Arthur Gelb
• Bayesian Multiple Target Tracking By Stone, Barlow & Corwin
• Multitarget-Multisensor Tracking : Principles & Techniques By
Yaakov Bar-Shalom & Xiao-Rong Li
• Multi-Sensor Fusion By Brooks and Iyengar