LERTEKNIK

REG

Lecture Outline
AU L
T O MA RO
T IC C O N T

Target Tracking: Lecture 6

Multiple Sensor Issues and Extended Target Tracking
Multiple Sensor Tracking Architectures
Emre Özkan Multiple Sensor Tracking Problems
[email protected]
Extended Target Tracking
Division of Automatic Control
Department of Electrical Engineering
Linköping University
Linköping, Sweden

December 17, 2014

E. Özkan Target Tracking December 17, 2014 1 / 53 E. Özkan Target Tracking December 17, 2014 2 / 53

Multi Sensor Architectures cont’d Multi Sensor Architectures cont’d

Figures taken from: M.E. Liggins and Kuo-Chu Chang

“Distributed Fusion Architectures, Algorithms, and Performance within a Network-Centric Architecture,”
Ch.17, Handbook of Multisensor Data Fusion: Theory and Practice, Taylor & Francis, Second Edition, 2009.

Figures taken from: M.E. Liggins and Kuo-Chu Chang

“Distributed Fusion Architectures, Algorithms, and Performance within a Network-Centric Architecture,”
Ch.17, Handbook of Multisensor Data Fusion: Theory and Practice, Taylor & Francis, Second Edition, 2009.

E. Özkan Target Tracking December 17, 2014 3 / 53 E. Özkan Target Tracking December 17, 2014 4 / 53
Multi Sensor Architectures: Centralized Multi Sensor Architectures: Hierarchical

Figures taken from: M.E. Liggins and Kuo-Chu Chang

“Distributed Fusion Architectures, Algorithms, and Performance within a Network-Centric Architecture,”
Ch.17, Handbook of Multisensor Data Fusion: Theory and Practice, Taylor & Francis, Second Edition, 2009.

Figures taken from: M.E. Liggins and Kuo-Chu Chang

“Distributed Fusion Architectures, Algorithms, and Performance within a Network-Centric Architecture,”
Ch.17, Handbook of Multisensor Data Fusion: Theory and Practice, Taylor & Francis, Second Edition, 2009.

E. Özkan Target Tracking December 17, 2014 5 / 53 E. Özkan Target Tracking December 17, 2014 6 / 53

Multi Sensor Architectures: Hierarchical with Feedback Multi Sensor Architectures: Decentralized

Figures taken from: M.E. Liggins and Kuo-Chu Chang Figures taken from: M.E. Liggins and Kuo-Chu Chang
“Distributed Fusion Architectures, Algorithms, and Performance within a Network-Centric Architecture,” “Distributed Fusion Architectures, Algorithms, and Performance within a Network-Centric Architecture,”
Ch.17, Handbook of Multisensor Data Fusion: Theory and Practice, Taylor & Francis, Second Edition, 2009. Ch.17, Handbook of Multisensor Data Fusion: Theory and Practice, Taylor & Francis, Second Edition, 2009.

E. Özkan Target Tracking December 17, 2014 7 / 53 E. Özkan Target Tracking December 17, 2014 8 / 53
Multi Sensor Architectures: Pros & Cons Problems in Multi Sensor TT

Registration: Coordinates (both time and space) of different sensors or

The traditional centralized architecture gives optimal performance but fusion agents must be aligned.
Bias: Even if the coordinate axis are aligned, due to the transformations,
• Requires high bandwidth communications.
biases can result. These have to be compensated.
• Requires powerful processing resources at the fusion center.
• There is a single point of failure and hence reliability is low. Correlation: Even if the sensors are independently collecting data, processed
For distributed architectures information to be fused can be correlated.
• Communications can be reduced significantly by communicating tracks Rumor propagation: The same information can travel in loops in the fusion
less often. network to produce fake information making the overall system
• Computational resources can be distributed to different nodes overconfident. This is actually a special case of correlation.
• Higher survivability.
• It is a necessity for legacy systems e.g. purchased radars might not Out of sequence measurements: Due to delayed communications between
supply raw data. local agents, sometimes measurements belonging to a target whose more
recent measurement has already been processed, might arrive to a fusion
center.

E. Özkan Target Tracking December 17, 2014 9 / 53 E. Özkan Target Tracking December 17, 2014 10 / 53

Correlation Correlation cont’d

Centralized Case Suppose the target follows the dynamics

Suppose the fusion center have the
yk1 prediction x̂k|k−1 in both cases. xk = Axk−1 + wk
Centralized Case
Fusion
x̂k|k and the ith sensor measurement is given as
Center  1     1 
yk C1 ek
= x k + yki = Ci xk + eik
yk2
yk2 C2 e2k

Decentralized Case where e1k and e2k are independent. Then with the KF equations
yk1 x̂1k|k
Tracker-1 Decentralized Case x̂ik|k =Ax̂ik−1|k−1 + Kki (yki − Ci Ax̂ik−1|k−1 )
=Ax̂ik−1|k−1 + Kki Ci (xk − Ax̂ik−1|k−1 ) + Kki eik
" #  
x̂1k|k
 1 
Fusion
x̂k|k I x̃k
Center = xk −
2
x̂k|k I x̃2k =Ax̂ik−1|k−1 + Kki Ci A(xk−1 − x̂ik−1|k−1 ) + Kki Ci wk + Kki eik
Tracker-2
yk2 x̂2k|k

E. Özkan Target Tracking December 17, 2014 11 / 53 E. Özkan Target Tracking December 17, 2014 12 / 53
Correlation cont’d Correlation cont’d
Define x̃ik , xk − x̂ik|k , then
Assuming that Σij0 = 0, we can calculate the correlation between the
x̃ik =xk − Ax̂ik−1|k−1
− Kki Ci A(xk−1 − x̂ik−1|k−1 ) − Kki Ci wk − Kki eik estimation errors of the local trackers recursively as
=Axk−1 + wk − Ax̂ik−1|k−1 − Kki Ci A(xk−1 − x̂ik−1|k−1 )
Σij i ij T j T i j
k = (I − Kk Ci )AΣk−1 A (I − Kk Cj ) + (I − Kk Ci )Q(I − Kk Cj )
T
− Kki Ci wk − Kki eik
=(I − Kki Ci )Ax̃ik−1 + (I − Kki Ci )wk − Kki eik
This necessitates that the fusion center knows the individual Kalman
Hence gains Kki and Kkj of the local trackers which is not very practical.

x̃ik =(I − Kki Ci )Ax̃ik−1 + (I − Kki Ci )wk − Kki eik Assuming that the errors are independent is not a good idea either.
Neglecting the correlation makes the resulting estimates overconfident
x̃jk =(I − Kkj Cj )Ax̃jk−1 + (I − Kkj Cj )wk − Kkj ejk
i.e., very small covariances meaning that too small gates and smaller
Kalman gains.
We can calculate the correlation matrix Σij i jT
k , E(x̃k x̃k ) as
When Q = 0, Σij
k = 0, i.e., no correlation when no process noise.
Σij i ij T j T i j
k = (I − Kk Ci )AΣk−1 A (I − Kk Cj ) + (I − Kk Ci )Q(I − Kk Cj )
T

E. Özkan Target Tracking December 17, 2014 13 / 53 E. Özkan Target Tracking December 17, 2014 14 / 53

Correlation Illustration Track Association: Testing

Maneuvers make this problem more dominant and visible. Test for Track Association [Bar-Shalom (1995)]:
120

Two estimates x̂ik|k , x̂jk|k and the covariances Σik|k , Σjk|k are given
σa = 0.3m/s2 100
2200

2000
from ith and jth local systems.
We calculate the difference vector ∆ij
80
1800
Position Errors (m)

1600
60
k
1400

∆ij j
y (m)

i
k , x̂k|k − x̂k|k
1200
40
1000

800
20

600

400

200
0
0 20 40 60 80
time (s)
100 120 140 160 Then we calculate covariance Γij ij ijT
k , E(∆k ∆k ) as
800 1000 1200 1400 1600 1800 2000 2200 2400 2600 2800
x (m) 120

σa = 0.3m/s2

Γij i j ij ijT
2200

k = Σk|k + Σk|k − Σk − Σk
2000 100

1800

80
1600

Then test statistics Dkij calculated as

Position Errors (m)

1400

60
y (m)

1200

1000

Dkij = ∆ijT ij −1 ij ij
40

k (Γk ) ∆k ≶ γk
800

600

20
400

can be used for checking track association.

200
800 1000 1200 1400 1600 1800 2000 2200 2400 2600 2800
x (m) 0
0 20 40 60 80 100 120 140 160
time (s)

E. Özkan Target Tracking December 17, 2014 15 / 53 E. Özkan Target Tracking December 17, 2014 16 / 53
 Track Association cont’d Track Association: Assignment Problem

What about the cross covariance Σij

k?
Simple method is to set it Σij
k = 0. Method proposed by [Blackman (1999)] for two local agents
It can be calculated using Kalman gains if they are transmitted to the fusion Form the assignment matrix:
center.
Approximation for cross covariance from [Bar-Shalom (1995)]: Aij T1j T2j T3j T4j NA1 NA2 NA3
T1i `11 `12 `13 `14 i
log βN A × ×
The following cross-covariance approximation was proposed:
 . 1 T2i `21 `22 `23 `24 × i
log βN A ×
Σij ≈ ρ Σ i
. ∗ Σj 2
T3i `31 `32 `33 `34 × × i
log βN A
k k|k k|k
βT Pi∈j Pj∈i N (x̂m
k|k
−x̂n
k|k
;0,Γmn
k )
where multiplication and power operations are to be done element-wise. For where `mn , log j .
βN A
negative numbers, square root must be taken on the absolute value and sign Then, use auction(Aij ) to get track association decisions.
must be kept.
The value of ρ must be adjusted experimentally. ρ = 0.4 was suggested for
2D tracking.

E. Özkan Target Tracking December 17, 2014 17 / 53 E. Özkan Target Tracking December 17, 2014 18 / 53

Track Association cont’d Track Fusion: Independence Assumption

Track association for more than two local agents. Once we associate two tracks, we have to fuse them to obtain a fused
track. This is called as track fusion.
One way is to solve multi dimensional assignment problem.
Consider the track fusion at point A assuming tCR = tCT = t.
The simpler way is to do the so-called sequential pairwise track
association.
Suppose we have NL local agents whose tracks need to be fused. Then, Independence assumption gives
we order the local agents according to some criteria e.g. accuracy, priority,
−1 −1 −1
etc. (ΣA
t ) =(ΣB
t ) + (ΣC
t )
−1 A B −1 B C −1 C
Tracks of Tracks of Tracks of
(ΣA
t ) x̂t =(Σt ) x̂t + (Σt ) x̂t
Local Agent 2 Local Agent 3 Local Agent NL

This is simplistic and expected to give

Track Track
very bad results here.
Tracks of Track Final
Association & Association &
Local Agent 1
Track Fusion
Association &
Track Fusion Track Fusion
Fused Tracks This is also called as naive fusion.

E. Özkan Target Tracking December 17, 2014 19 / 53 E. Özkan Target Tracking December 17, 2014 20 / 53
Track Fusion cont’d Extended Target Tracking
Illustration of Correlation Independent Schemes.
−1
z T (ΣB
t ) z=1
Single extended target modeling
• Target state: kinematics and shape
• Number of measurements per target
z T (ΣC
t )
−1
z=1 • Motion modeling not covered. Typically possible to use point target
models.

Multiple extended target tracking

• Data association for extended targets
−1
• Measurement set partitioning
z T (ΣA
t ) z=1
Largest Ellipsoid
Algorithm
Covariance Intersection

E. Özkan Target Tracking December 17, 2014 21 / 53 E. Özkan Target Tracking December 17, 2014 22 / 53

Extended Target Tracking Extended Target Tracking

One measurement per target is often not valid, e.g.,

laser sensors, camera images, or automotive radar. Multiple measurements per target ⇒ possible to estimate the target’s
12
extension.
10
• The spatial extension, i.e. size and shape.
8
• Hence the name extended target.
6

4 The extension estimate provides additional information, e.g. for

2 classification and identification of different target types:
0
−10 −5 0 5 10
• Human?
• Bike?
• Car?
This leads us to the following definition:
• etc...
Extended targets are targets that potentially give rise
to more than one measurement per time step.

E. Özkan Target Tracking December 17, 2014 23 / 53 E. Özkan Target Tracking December 17, 2014 24 / 53
Group targets Single extended target

Airplane tracked with radar

Closely related to extended targets is group targets.
Reflection points, or sources, 4
spread across surface. Locations
and number vary with time.
Closely spaced point targets Measurements zk may fall
“outside” target due to noise
Cannot be resolved Extended target state ξk
contains parameters for
Treated as single group target • Kinematics
• Extension

Not interested in sources, want to estimate the state ξk given the sets of
measurements
 
p ξk |Zk , Zk = z1k , . . . , zm Zk = {Z1 , . . . , Zk }

Presented extended target models applicable to group targets. k
k
,

E. Özkan Target Tracking December 17, 2014 25 / 53 E. Özkan Target Tracking December 17, 2014 26 / 53

Target shape – single ellipse Target shape – single ellipse

Approximate shape as a single ellipse

Vector xk and positive semi-definite shape matrix Xk , “Extended Object and Group Tracking with Elliptic Random Hypersurface Models”, Baum et al.

(zk − xk )T Xk−1 (zk − xk ) ≤ 1 Cholesky decomposition of shape matrix Xk = Lk LTk

 
ak 0
Lk =
bk ck
T
Extended state vector ξk = xTk ak bk ck


Assumed Gaussian models

 
p ξk |Zk =N ξk ; mk|k , Pk|k

p (zk |ξk ) =N (zk ; hk (ξk ) , Rk )

E. Özkan Target Tracking December 17, 2014 27 / 53 E. Özkan Target Tracking December 17, 2014 28 / 53
Target shape – single ellipse Target shape – single ellipse

“Extended Object and Group Tracking with Elliptic Random Hypersurface Models”, Baum et al. “Bayesian Approach to Extended Object and Cluster Tracking using Random Matrices”, Koch.

State ξk is combination of kinematical vector xk and extension matrix

Simulation results (plots copied from paper)
Xk , ξk = (xk , Xk ). Gaussian inverse Wishart distributed,
 
p ξk |Zk = N xk ; mk|k , Pk|k ⊗ Xk IW Xk ; vk|k , Vk|k

Called random matrix model

Measurement model

p (zk |ξk ) = N (zk ; Hk xk , Xk ) ,

i.e. the covariance is given by the extension matrix.

E. Özkan Target Tracking December 17, 2014 29 / 53 E. Özkan Target Tracking December 17, 2014 30 / 53

Target shape – single ellipse A simple random matrix model.

“Bayesian Approach to Extended Object and Cluster Tracking using Random Matrices”, Koch. State ξk is combination of kinematical vector xk and extension matrix
Pros: Linear measurement update equations for extension state. Xk , ξk = (xk , Xk ). Gaussian inverse Wishart distributed,
Cons: Difficult to model sensor noise. The measurement model
 
p ξk |Zk = N xk ; x̂k|k , Pk|k IW Xk ; vk|k , Vk|k

p (zk |ξk ) = N (zk ; Hk xk , Xk )

where,
implies that sensor noise is much smaller than extension. v−d−1 v−d−1
2− 2 d |V |− 2 1
IW (X ; v, V ) = v exp tr(− X −1 V )
Γd ( v−d−1
2 )|X| 2 2
“Tracking of extended objects and group targets using random matrices”, Feldmann et al.
Measurement model
Alternative measurement model
p (zk |ξk ) = N (zk ; Hk xk , Xk ) ,
p (zk |ξk ) = N (zk ; Hk xk , sXk + Rk )
i.e. the covariance is given by the extension matrix.

E. Özkan Target Tracking December 17, 2014 31 / 53 E. Özkan Target Tracking December 17, 2014 32 / 53
A simple random matrix model. A simple random matrix model.

Prediction Update

x̂k+1|k = Ax̂k|k , Measurement Update

Pk+1|k = APk|k AT + BQB,
   
p xk , Xk |Zk ∝ p(Zk |xk , Xk )p xk , Xk |Zk−1
vk+1|k = λvk|k ,
Vk+1|k = λVk|k , Inverse Wishart defines a conjugate prior for the Gaussian likelihood.
A family of prior distributions is conjugate to a particular likelihood
where 0.5 < λ < 1 is the forgetting factor. function if the posterior distribution belongs to the same family as the
Measurement model prior.
Hence the posterior is in the same form.
p (zk |ξk ) = N (zk ; Hk xk , Xk ) ,

i.e. the covariance is given by the extension matrix.

E. Özkan Target Tracking December 17, 2014 33 / 53 E. Özkan Target Tracking December 17, 2014 34 / 53

A simple random matrix model. A simple random matrix model.

where,
Measurement Update mk
1 X
• Kinematic State z̄k = zki ,
mk
i=1
x̂k|k = x̂k|k−1 + Kk (z̄k − C x̂k|k−1 ), mk
X
Pk|k = Pk|k−1 − Kk Sk KkT , Z̄k = (zki − z̄k )(zki − z̄k )T ,
i=1
where, Vk|k−1
X̂k|k−1 =
X̂k|k−1 vk|k−1 − 2d − 2
Sk = CPk|k−1 C T + , Kk = Pk|k−1 C T Sk−1
m “Tracking of extended objects and group targets using random matrices”, Feldmann et al.
• Extent State State of the art filter
vk|k = vk|k + m,
p (zk |ξk ) = N (zk ; Hk xk , sXk + Rk )
Vk|k = Vk|k + Z̄k ,
Better sensor noise model, but the measurement update is heuristic
and no longer linear.
E. Özkan Target Tracking December 17, 2014 35 / 53 E. Özkan Target Tracking December 17, 2014 36 / 53
Target shape – single ellipse Target shape – multiple ellipses

A single ellipse can be a crude model, multiple ellipses gives better

Random matrix model popular in the literature, see e.g. approximation, see e.g.
• “Probabilistic tracking of multiple extended targets using random matrices”, • Arbitrary number of ellipses, not same center point.
Wieneke and Koch. “Tracking of Extended Object or Target Group Using Random Matrix – Part II:
• “A PHD filter for tracking multiple extended targets using random matrices”,
Granström and Orguner. Irregular Object”, Lan and Rong Li.
• “Tracking of Extended Object or Target Group Using Random Matrix – Part I: New
Model and Approach”, Lan and Rong Li.
and the reference therein.

Single ellipse also used by Degerman et al in “Extended Target

Tracking using Principal Components”.
• The shape matrix decomposed into rotation, major axis, and minor axis.

E. Özkan Target Tracking December 17, 2014 37 / 53 E. Özkan Target Tracking December 17, 2014 38 / 53

Taregt shape – Star convex Target shape – general shapes

A set S(xk ) is called star-convex if each line segment from the center
to any point is fully contained in S(xk ). Single, or multiple, ellipses can be a crude model of the shape. Need
for more general shape model.
y-coordinate

1 “Shape Tracking of Extended Objects and Group Targets with Star-Convex RHMs”,
r
0 θ Baum and Hanebeck.
The shape is given by a radial function r(θ) – distance from center of
−1 target to edge of shape at angle θ.
−1 0 1 • Circle: r = f (θ) p
x-coordinate • Canonical ellipse: r(θ) = a2 cos(θ)2 + b2 sin(θ)2
• General case: truncated Fourier series expansion
FN
Radius r

a0 X
1 r(θ) = + aj cos(jθ) + bj sin(jθ)
2 j=1
r = f (θ)
0 Shape parameter: p = [a0 , a1 , . . . , aNF , b1 , . . . , bNF ]T
0 π 2π
Angle θ
E. Özkan Target Tracking December 17, 2014 39 / 53 E. Özkan Target Tracking December 17, 2014 40 / 53

Figure: Description of star-convex shapes using a radius function r = f (θ).

Target shape – general shapes Multiple extended target tracking

“Shape Tracking of Extended Objects and Group Targets with Star-Convex RHMs”,
Assume that we have chosen the random matrix model for the
Baum and Hanebeck.
extended targets.
Compare with ellipse (plots copied from paper):
How can we handle multiple extended targets, when there is clutter
and detection uncertainty?
As with point target tracking, the data association problem must be
solved.
Different because each target can generate multiple measurements.
• Which measurements were caused by clutter sources?
• Which measurements were caused by targets? Of those, which were
caused by the same target?

E. Özkan Target Tracking December 17, 2014 41 / 53 E. Özkan Target Tracking December 17, 2014 42 / 53

Data association for extended targets Partitioning the measurements — example

n o
(1) (2) (3)
Partition the measurement set Zk = zk , zk , zk

Z p1 p2
1.5 1.5 1.5

One approach: Partitioning of the measurement set 1 (2)

zk
(1)
zk 1 1
W12
W11
0.5 0.5 0.5

y
0 (3)
zk 0 0 W22
A partition p is a division of the set Zk into non-empty subsets, called −0.5 −0.5 −0.5
−0.5 0 0.5 1 1.5 −0.5 0 0.5 1 1.5 −0.5 0 0.5 1 1.5
cells W . x x x

p3 p4 p5
1.5 1.5 1.5

Interpretation: One cell = one source. 1 1 1 W25

y W23 W14 W15

y
0.5 0.5 0.5
W13 W24
0 0 0 W35

−0.5 −0.5 −0.5

−0.5 0 0.5 1 1.5 −0.5 0 0.5 1 1.5 −0.5 0 0.5 1 1.5
x x x

E. Özkan Target Tracking December 17, 2014 43 / 53 E. Özkan Target Tracking December 17, 2014 44 / 53
Partitioning method Partitioning method – laser example

Intuition: Measurements are from same source if they are close, with
For optimality we must consider all possible partitions.
respect to some measure or distance.
Number of possible partitions for n measurements
given by n:th Bell number Bn .
12

The Bell numbers Bn increase very fast when n increases,

10

e.g. B3 = 5, B5 = 52 and B10 = 115975.

8

Computationally infeasible to consider all partitions. 6

Necessary to approximate the full set of partitions 2

with a subset of partitions. 0

−10 −5 0 5 10

E. Özkan Target Tracking December 17, 2014 45 / 53 E. Özkan Target Tracking December 17, 2014 46 / 53

Partitioning method – laser example Partitioning method

Method: Measurements are in same cell W if distance is “small”.

Intuition: Measurements are from same source if they are close, with
respect to some measure or distance. Distance between two cells Wj and Wi is measured as

8 min {d (zi , zj ) : zi ∈ Wi , zj ∈ Wj }
7

12
6
Partitions pi where all pairwise cell distances is < di .
10

8
5 Limit to partitions for thresholds di that satisfy
4

dmin ≤ di < dmax

6

3
4

2 2
If possible, use scenario knowledge to choose
0
−10 −5 0 5 10
1
−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 distance measure and to determine bounds.
Important to choose dmin and dmax conservatively.

E. Özkan Target Tracking December 17, 2014 46 / 53 E. Özkan Target Tracking December 17, 2014 47 / 53
Partitioning example Extended Target Models
p1 = W11 , W21 , W31 p2 = W12 , W22 , W32 , W42
 

8 8

7 7

6 6
Extended target models applicable in scenarios where there are
5 5
multiple measurements per target.
4 4

3 3 Also applicable when multiple targets move in tight groups.

2 2

1
−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3
1
−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 Shape models range from simple ellipse to general shapes.

Reasonable to discard most partitions as highly unlikely. Multiple extended target tracking possible using, e.g., the extended
Additional methods given in: target PHD filter.
“Extended Target Tracking using a Gaussian-Mixture PHD filter”, Granström et al.
“A PHD filter for tracking multiple extended targets using random matrices”, Granström
and Orguner

E. Özkan Target Tracking December 17, 2014 48 / 53 E. Özkan Target Tracking December 17, 2014 49 / 53

Extended Multi-target Tracking- Recent Works Extended Target Examples

Laser Pedestrian
Video Pedestrian
Laser - Mix
“A PHD filter for tracking multiple extended targets using random matrices”, Granström
and Orguner.
30

“A Multiple-detection joint probabilistic data association filter”, Habtemariam et al. 25

20
“A multiple hypothesis tracker for multitarget tracking with multiple simultaneous 14

measurements”, Sathyan et al.

y [m]

15 12

10

y [m]
10
8

6
5

0
−20 −15 −10 −5 0 5 10 15 20 −20 −15 −10 −5 0
x [m] x [m]

E. Özkan Target Tracking December 17, 2014 50 / 53 E. Özkan Target Tracking December 17, 2014 51 / 53
References References

M. E. Liggins II, Chee-Yee Chong, I. Kadar, M. G. Alford, V. Vannicola and S. Blackman and R. Popoli, Design and Analysis of Modern Tracking Systems.
V. Thomopoulos, “Distributed fusion architectures and algorithms for target tracking,” Norwood, MA: Artech House, 1999.
Proceedings of the IEEE, vol.85, no.1, pp. 95-107, Jan. 1997.
Y. Bar-Shalom and X. R. Li, Multitarget-Multisensor Tracking: Principles, Techniques.
M. E. Liggins and Kuo-Chu Chang, “Distributed fusion architectures, algorithms, and Storrs, CT: YBS Publishing, 1995.
performance within a network-centric architecture,” Ch.17, Handbook of Multisensor
Data Fusion: Theory and Practice, Taylor & Francis, Second Edition, 2009. Granström et al., Extended Target Tracking using a Gaussian-Mixture PHD filter

K. C. Chang, Chee-Yee Chong and S. Mori, “On scalable distributed sensor fusion,” Granström and Orguner, A PHD filter for tracking multiple extended targets using
Proceedings of 11th International Conference on Information Fusion, Jul. 2008. random matrices

E. Özkan Target Tracking December 17, 2014 52 / 53 E. Özkan Target Tracking December 17, 2014 53 / 53