Downloaded from [Link]

ir
[Link] [Link]

19th International Conference on Information Fusion

Heidelberg, Germany - July 5-8, 2016

Multiple Sensor Bayesian Extended Target Tracking

Fusion Approaches Using Random Matrices
Gemine Vivone*, Karl Granstrom t, Paolo Braca* and Peter Willett+
*NATO STO Centre for Maritime Research and Experimentation
19126 La Spezia, Italy
Em~~: {[Link],[Link]}@[Link].
tDepartment of Signals and Systems, Chalmers University of Technology
41296 Gothenburg, Sweden
Email: [Link]@[Link].
+Department of Electrical and Computer Engineering, University of Connecticut
Storrs, CT 06269-2157 USA
Email: willett@engr .ucon n .edu.

Abstract-The tracking of extended targets is attracting a of the random matrices into the probabilistic multi-hypothesis
growing literature thanks to the high resolution of several modern tracking and into the PHD and CPHD filters, respectively, is
radar systems. A fully Bayesian solution has been proposed addressed for the multi-target tracking problem. Furthermore,
in the random matrix framework. In this paper, the fusion
of detections acquired by multiple sensors is analyzed. Four a new approach is derived in [7] to overcome some of the
different methods are proposed to track and to estimate jointly weaknesses in [14] . Indeed, in [14] sensor inaccuracies are
both the kinematic and extent parameters. All of them use neglected and, if they are large in comparison to the target
the same multi-sensor kinematic vector measurement update. size, the lack of modeling may lead to an overestimation
The first approach is based on a particle approximation of the of the target size, see [6] . New measurement and time up-
extent state probability density function, whereas the other three
are based on an inverse Wishart representation of the latter. dates for [7] are proposed in [20] and [12], respectively. An
Extensive simulations evaluate the performance of the different extension of random matrices for non-ellipsoidal group and
approaches. The best performance is obtained by the particle extended target tracking based on a combination of multiple
filter-based approach paid by an increased computational burden. ellipsoidal sub-objects, each represented by a random matrix,
Comparable performance are observed for the two updates based is discussed in [16]. An interesting application of the random
on multi-sensor generalization, while the worst performance is
obtained by the updated based on fusion approximation. matrix framework is reported in [8], [9] using radar data. The
problem of the conversion between polar/bistatic and Cartesian
I. INTRODUCTION coordinates into random matrices is addressed in [21], [22],
High resolution radars are becoming very common and respectively, and tested on real data.
widespread. In such systems targets appear as a cloud of de- The original single sensor random matrix model, presented
tections instead of a point. Standard target tracking algorithms in [14], modeled the dependence between the kinematic state
rely upon the hypothesis of at most one detection per target, and the extent state. Later work showed that by not explicitly
which is no longer valid in this case. Therefore, the develop- modeling this dependence, more general models for both
ment of new strategies for target tracking is required to deal sensor noise and dynamic target motion could be used [7],
with this issue. The rubric is extended target tracking (ETT) [11]. Results using both simulated and experimental data
in the related literature. showed that the single target updates and predictions provide
Several approaches have been already proposed, see e.g. [3], for sufficient interdependence between the kinematic state
[4], [18]. An overview of the state-of-art for group and and the extent state to compensate for lack of an explicit
extended target tracking techniques is given in [19] . A pop- model of the dependence. Following the related literature, the
ular and computationally efficient framework for extended explicit modeling of the dependence between the kinematic
target tracking is provided by Koch in [14]. An approximate state and the extent state is neglected here. Thus, in this
Bayesian solution to the target tracking problem is illustrated paper, the multiple sensor fusion problem is solved for the
under the hypothesis of an elliptical spread of the target. kinematic update in a closed form and four different methods
Random matrices are exploited to model the ellipsoidal object are proposed to update the extent state and compared using
extensions, which are treated as additional state variables to be extensive simulations.
estimated or tracked. The target kinematic states are modeled The paper is organized as follows. The paper's contributions
using a Gaussian distribution, whereas the ellipsoidal target are explained in Sect. II. The four extensions to the multiple
extension is modeled using an inverse Wishart distribution. sensor case are illustrated in Sect. III. The extended target pre-
Random matrices are used to model extended targets under diction is reviewed in Sect. IV. Numerical results are presented
kinematic constraints in [15]. In [10], [17], [23], the integration in Sect. V. Finally, conclusions are drawn in Sect. VI.

978-0-9964527 -4-8©2016 ISIF

Authorized licensed use limited to: ULB Darmstadt. Downloaded on May 22,2025 at 20:16:33 UTC from IEEE Xplore. Restrictions apply.
Downloaded from [Link]
[Link] [Link]

II. PAPER CONTRIBUTIONS For the extent state, two different approximations will be
considered. The first is a particle approximation of the extent
Consider a multistatic radar system that consists of S pdf,
sensors that acquire data within the same surveillance area.
The set of measurements acquired at time k by the s-th sensor Np

is denoted P ( Xk zIc,k- 1) ~ "" W(i)

~ ~ klk- 16" X k. (x(i))
k (8)
i= l
Zs = {ZS,j }n~ Np
~ L Wkl~Oxk (Xki))
(1)
k k j =l P (Xk IZC,k) (9)
i= l
The detections z~,j are assumed independent and each sensor
s acquires nic measurements of the target at time k. The full Here the paper presents the following contribution,
set of detections at time k is the union of the sets of detections • An update based on a Rao-Blackwellized particle repre-
from the S sensors, sentation of the extended target pdf.
s The second extent state pdf approximation that is considered
Zk = U Zic (2) is an inverse Wishart pdf,
s= l
p(Xk IZC ,k- 1) ~IWd (Xk; vk lk- 1, Vklk- 1) (lO)
Let z c,k = {ZR} ;=0
denote all such measurement sets up to, p(Xk IZC, k) ~IWd (Xk; vk lkl Vk lk ) (11)
and including, time k.
The multi-sensor measurement likelihood, conditioned on where the inverse Wishart pdf is defined as in Table I. Here
the extended target state (which consists of the target kinematic the paper presents the following three contributions:
state Xk and Xk that describes the target's extent exploiting • An update based on approximate fusion.
an elliptical model) and the number of detections nic, is • A multi-sensor generalization of the single sensor update
from [7J.
P (Zk IXk' Xk, {nk} ;=1) • A multi-sensor generalization of the single sensor update
s nk
from [IJ.
= II II N (z~,j ; HkXk, pXk + Ric) (3) The four different updates are compared in an extensive sim-
ulation study in which a network of radar sensor is simulated.
s= l j = l
Some notation that is used throughout the paper is listed in
where p is a scaling factor, H k is a measurement model at Table I.
time k that selects the position components in the state vector,
and Ric is the covariance matrix for the sensor s at time k. III. MULTI-S E NSOR UPDATES
Remark: For brevity and notational simplicity, the mea-
surement noise covariance's (RV possible dependence on This section presents the kinematic state update and the four
the kinematic state is not explicitly indicated. However, all different multi-sensor updates.
derivations in the paper are valid also in the case of kinematic
state dependent covariance matrices, see [21]. A. Gaussian kinematic state update
Given a prior distribution for the extended target state,
The kinematic state is updated as follows
(4)
m~~ =mk lk- 1 + K k1k- 1 ( zk lk- 1 - Hkmk lk- 1) (18)

the objective of the paper is to find the posterior P~~ =Pklk- 1 - K k1k- 1HkPk lk- 1 (19)

(5) where

Y~k-1 = ~ ( Yklk- 1) -1)

S - 1
These densities approximate the kinematic and extent states as (
(20a)
independent, however, as noted in [7], the measurement update
step "provides for the interdependency between kinematics and
extension estimation" . Sk lk- 1 =Hk Pklk- 1H 'k + Ykclk_1 (20b)
The prior and posterior kinematic state probability density K k1k- 1 =Pklk- 1H'k (Sk lk_1) - 1 (20c)
functions (pdfs) are approximated by Gaussian distributions,

p(xk IZC,k- 1) ~N (Xk; mk lk- 1, Pklk- 1) (6) zk lk- 1 =Y~k-1 {t, [(Yk1k- 1) - 1 Zic] } (20d)

p(xk IZC,k) ~N (Xk; mk lkl Pk lk ) (7)

Authorized licensed use limited to: ULB Darmstadt. Downloaded on May 22,2025 at 20:16:33 UTC from IEEE Xplore. Restrictions apply.
Downloaded from [Link]
[Link] [Link]

TABLE I value. In other words, we approximate Xk ~ mklk- 1, and the

NOTATIONS single sensor log-likelihood can be written as
• IRn is the set of real column vectors of length n, is the set of §++ IXk, X k, nk)
symmetric positive definite n x n matrices, §+
is the set of sy mmetric positive log p (Zk (23)
semi-definite n x n matrices, and N is the set of non-negative integers.
• N (x i m, P) denotes a multi-variate Gaussian pdf over the vector x E nk
~ - 2log det (pXk k - "21 tr
+ R 8) [(pXk + Rk8) -1 ~k8]
IRn x with mean vector m E IRn x , and covariance matrix P E §~x .
• IW d (X i v , V) denotes an inverse Wishart pdf over the matrix X E §i+ where
with scalar degrees of freedom v > 2d and parameter matrix V E §i+ , see, n~
e.g. [13, Definition 3.4.1]
~k = L (z~,j - Hkmk lk- 1) (z~,j - Hkmk lk- 1) (24)
(1
T
v - d- l v - d- l
T - 2 - det(V)-2- - 1 ) j=l
IW d( X i V, V)= ( ) et r - -X V , (12)
= Zk+ nk(zk - Hk m klk- 1)(zk - Hkmk lk- 1r
v
I'd V-~- l det( X )2 2 (25)

where etr(-) = exp(tr(-)) is exponential of the matrix trace, det (-) is the and all terms that are independent of the detections and the
determinant operator, and I'd (.) is the multivariate gamma function. The target state are omitted.
multivariate gamma function I' d (.) can be expressed as a product of the The logarithm of the multi-sensor likelihood (3) is the sum
ordinary gamma function I'(-) , see [13 , Theorem 1.4.1].
• W d(X i w, W ) denotes a Wishart pdf over the matrix X E §i+ with of the single sensor log-likelihoods (23),
scalar degrees of freedom W ~ d and parameter matri x W E §i+, see, e.g. s
[13 , Definition 3.2.1]
log p (Z k IXk, X k, {nk} ;=1 ) = L logp (Z k IXk, Xk, nk)
T v~d det (X ) W - ; - l ( ) 8=1
W d( X i W, W) = etr --21 W - 1X . (13) (26)
I' d (~) det(W) '1
s
• Expected prior extent [13]
~ -~ L n k logdet(pXk + R k ) + tr [(pXk + Rk) -l~kJ
~ V k1k - 1 8=1
X kIk- 1 = _ _.:.:..c..:--c::"-,-----,- (14)
Vklk- 1 - 2d - 2 (27)
• Expected added effect of extent and sensor noise, normalized by number = log Lk(Zk IXk) (28)
of detections,
where Lk(Zk IXk) is shorthand notation for the approximate
(15) multi-sensor likelihood as a function of the set of measure-
ments Zk and the extent state Xk. The particle filter (PF)
• Centroid measurement and measurement spread for sth sensor are proceeds to approximate the posterior at time k by a new
1
ZS = _ '"""
n S
k .
Zs - '\""'
n S
k
(ZS,j _ zs) (ZS,).
_ zs) T set of weighted particles {Wki) ,Xki)}
~ as described in the
Z S,) . (16)
k s
nk
6
j=l
k ' k- ~ k k k k sampling importance resampling (SIR) fifter [2].
j=l
Particles are sampled from the importance sampling distri-
Note: measurement spread is not equivalent to sample covariance.
bution that is represented by a Gaussian distribution along the
• Cholesky factorization of positive definite matri x A is denoted
ellipse orientation, minor axis, and major axis extracted from
T 1
A = A2 Az (17) the particle Xk~ 1.
Then particle weights are updated based on the likelihood
function of the observed data Zk from all sensors . The
likelihood Lk (Zk IXk) is given in Eq. (28) , and takes into
B. Particle approximation extent update account all information at time k. Lastly, a resampling strategy
is adopted to avoid the particle degeneracy problem, e.g.
The Bayesian extent state posterior is usually impossible to see [2].
evaluate in closed form. Different approaches to approximating A target extent estimate is given by the Minimum Mean
it exist, e.g., Sequential Monte Carlo (SMC) methods [2], Square Error (MMS E) estimator, which is optimal in terms of
[5] . Assume that at time k - 1, a set of weighted particles MSE. Afterwards, the estimation at time k, i.e. Xk lk> is used
representing the posterior is available for the next prediction step.

{ Wk-
(i)
1,
(i)
X k_ 1
}Np
. (21)
C. Inverse Wishart extent updates
,=1 1) FFK generalization: The multi-sensor generalization of
the FFK (here called FFK, after the authors initials) single
(22) sensor update [7] is

(29)
where Oy (X) is a Dirac delta function centered in y. 8=1
Similarly to how in the kinematic state update the extent S
state is approximated by its predicted expected value, here we Vk~ksFFK =Vklk- 1 + L Zklk-1 + N k1k - 1 (30)
approximate the kinematic state with its predicted expected 8=1

Authorized licensed use limited to: ULB Darmstadt. Downloaded on May 22,2025 at 20:16:33 UTC from IEEE Xplore. Restrictions apply.
Downloaded from [Link]
[Link] [Link]

where Y [m]
T 1

Zklk- 1 =X~k_ 1 (Yk1k- 1) - 2 Zk (Yklk_1f2 X!lk_1

(31) S4 ~ S3
T (0,5000) ----------------~

Nk1k- 1 =X~k_ 1 (Sk lk- 1) - (Zklk- 1 - Hkmk lk- 1)

r
2 Target
1

X (Zk lk-1 - Hkmk lk- 1 (Sk lk- 1) - 2 X!lk_ 1 Random target

trajectory
(32)
2) ULL generalization: The multi-sensor generalization of
the ULL (unbiased log-likelihood Iinearisation) single sensor
update [1] is
(0,0) (5000, 0) X [m]
S
MS ULL
v klk = vk lk- 1 + '"
~ nkS (33)
s= l Fig. 1. Simulator setup with four sensors and random target trajectories.
S
[Link]
MS ULL
= v klk- 1 + ~
TT ' " 1\J S
'lk lk-1 (34)
s= l where <1?O is a rotation matrix and the Wishart pdf is
where defined as in Table I. Given the posterior extent Xk and the
posterior kinematic state Xk, the expected predicted extent is
M k1 k- 1 = n k X klk- 1 + n kPXk lk- 1(Cklk_ 1) - 1 <1?(Xk)Xk<1?(Xk)T. The parameter Vk+1,k governs the extent
process noise: the larger Vk +1 ,k is, the lower the process noise.
x (W k1k- 1- Cklk- 1) (Cklk_1)-lXklk- 1 (35)
Thus, the extent prediction corresponds to rotating the extent
1 n~ . by an amount specified by the kinematic state, and increasing
L
.
- Hkmk lk- 1) (z~'J - Hkmk lk- 1)
T
Wk1k- 1 = n S (z~'J the variance. In [11] the predicted parameters vk+ 1lk and
k j=l Vk+1l k are computed using a series of density approximations
(36) that minimize the Kullback-Leibler divergence. Here we use
(37) a computationally efficient approximation and update the
parameters as follows
3) Fusion approximation: This update is based on fusion
approximation (FA) and was presented in [22]. Vk+1l k =2d + 2 + e- T / T (vk lk - 2d - 2) (43)
<1?(xklk)Vk lk<1?(Xk lk)
IV. EXTENDED TARGET PREDICTION Vk+1 lk = ( vk+ 1lk - 2d - 2) 2d 2 (44)
vk lk - -
For completeness, the extended target prediction is given
where T is a time constant related to the agility with which
here. The kinematic state transition density is
the target may change its extent over time. Thus, in this
approximate prediction the parameter T replaces the Wishart
transition density degrees of freedom Vk +1, k. However, note
where f k+ 1,k (Xk) is a function that describes the motion of that the effect of the prediction is the same as if the prediction
the target. The updated mean and covariance are given by the from [11] had been used: the extent expected value is rotated,
extended Kalman filter prediction and the covariance is increased.
mk+1l k =fk+1 ,k(Xk) (39) V. NUMERICAL SIMULATION STUDY
Pk+1 lk =Fk+1, k P klk F'{+l ,k + Qk+1 (40) The presented Bayesian extended target tracking updates are
where compared using simulated data.
A. Setup
The simulator setup is shown in Figure 1. The target's
is the Jacobian of f k+1 ,k(Xk) with respect to the state Xb initial positions are randomly sampled from inside the sensor
evaluated at the expected value xk lk. network, and the target tracks are randomly generated. The
In the case the extent pdf is approximated by an inverse simulated target follows both linear and curved trajectories.
Wishart pdf, the extent state transition density is The number of detections per scan is variable and depends on
the distance between the target and the sensor, as well as on
the simulator's parameters, i.e., the azimuthal accuracy and the
(42) range resolution. The parameters are summarized in Table III.
The goal of this simulator is to obtain average performance of

Authorized licensed use limited to: ULB Darmstadt. Downloaded on May 22,2025 at 20:16:33 UTC from IEEE Xplore. Restrictions apply.
Downloaded from [Link]
[Link] [Link]

2.5 10
- PF
I-- MSFFK
PF
MSULL
- FA

;;; 7
E
- 6
Z- ~
c;,
"g 5 §
~
4
is
"
W 3

0
0 50 100 150 200 50 100 150 200
Time (5] Time (5]

(a) (b) (c)

- PF
- MSFFK
- MSULL
- FA
,;,
~ 30
~
15
.~

1000" - - -50
= -- -.,.:-
1000 --------c
,500-
0 - - - - - 0200
' lO'0"---50
=---.,.:-
1 000--------c,50
:O---~
2oo
Time (s] Time (s]

(d) (e) (f)

Fig. 2. Errors in (a) position, (b) velocity, (c) length, (d) width, (e) orientation , and the (f) Frobenius matrix error are depicted for the four compared
approaches averaged on 103 Monte Carlo trials on Simulator 1.

TABLE II
ESTIMATION ERRORS AVERAGED ON 103 MONTE CARLO TRIALS

Algorithm EP oS [m] Ev el [m·s I] EW2 a [m] E1en [m] Eor [deg.] EFE

FA 1.75 0.328 49.1 21.3 18.1 1698.7

MSULL 1.88 0.324 30.0 10.9 14.0 969.3
MSFFK 1.95 0.324 22.2 10.3 16.3 830.1
PF 2.l7 0.317 3.2 2.0 1.72 110.9

the different updates when Monte Carlo trials are performed. rotation matrix is defined as
Table IV shows the sets of parameters used by the compared
updates. 1>(Xk) = [C~S(WkTs ) - Sin(WkTS)] (46)
sm(wkTs) COS(WkTs)
The random vector xk = [Pb vk, Wk r E jR5 is the
kinematical state, and describes the target's position Pk E jR2, The performance metrics used for the assessment are the
velocity Vk E jR2 and turn-rate Wk E JRI. The motion model root mean square errors in position (i.e. ePOS), velocity (i.e.
fC) and process noise covariance Q are ev el ), length (i.e. elen ), width (i.e. ewid ), and orientation (i.e.
eaT). Furthermore, the Frobenius matrix error (i.e. eF E ) for
s in(w kTs ) - 1+cos(w kTs )
1 0 Wk
0 the target extent Xk is adopted to have an overall performance
1 l -co~(W k Ts ) sin(wk T s )
0 Wk Wk
0 index for the target's extent estimation.
f(Xk) = 0 0 COS(WkTs) - sin(wkTs) 0 Xk,
B. Results
0 0 sin(wkTs) COS(WkTs) 0
0 0 0 0 1 Figure 2 shows the results averaged over 103 Monte Carlo
(45a) trials. The PF clearly has the smallest extent estimation errors

Q = Gdiag ( [O'~ , 0'2

a , O'~]) G T , G =
[T;I
fsI:
0,,,] 02XI
and also converges faster, however, the price for the improved
performance is an increase in the computational burden. How
the computational time scales with the number of particles and
Olx2 Ts the related improvements in the Frobenius matrix error metric
(45b) are shown in Figure 3. Thus, a number of particles of 103 (used
where Ts is the sampling time, 0' a is the acceleration standard in this paper) can be advisable given the trade-off between ac-
deviation, and O'w is the turn-rate standard deviation. The curacy and computational time. The other approaches require

Authorized licensed use limited to: ULB Darmstadt. Downloaded on May 22,2025 at 20:16:33 UTC from IEEE Xplore. Restrictions apply.
Downloaded from [Link]
[Link] [Link]

~6
c
ro
1000
if> 5
Qi
E a.
0
c
Q) 4
E
x i=
' ':::
ro
:2
~ 3
o
500 ~
S 2
a.
E
o
01

100 199 398 794 1583 3158 6300 1256625066

No. Particles

(a) (b)

Fig. 3. The hi stogram representation of the Frobenius matri x error versus the number of particles is depicted in (a), whereas the computational time per scan
versus the number of particles for the particle filter-based approach is shown in (b).

TABLE III X 104

SIMUL ATOR PA RA METER SETTING 3.5 i - - - - - - - , - - - - - - , - -r===::=::==il
- PF
Parameter Value Specification -MSFFK
3 . . ... - MSULL
Ts 1s Sampling time - FA
O"pos 0.1 m . s- 2 Std. process noise
Maximum turn-rate 2.5
Bm ax 1°
(T ¢ 0.1 ° Std. turn-rate noise
(T h 1° Std. heading noise ~ 2
It 70 Target length C,
Wt 15 Target width .9
~ 1.5
k max 200 Number of frames
~r 0.5 m Range resolution
hbw 0.1 ° Azimuthal accuracy
V max 15 Maximum velocity target
Vmin 0.5 Minimum velocity target
Nd variable Num. detects. time per sensor
S 4 Num. of sensors
(Tr 10m Std. noise r ange 1000 1500
(Te 2° Std. noise azimuth Matrix Norm

Fig. 4. Histogram of the Frobenius matrix errors for the four compared
TABLE IV
approaches averaged on 10 3 Monte Carlo trials.
TR ACKING PA RA METER SETTI NGS

Parameter Value Specification

Ts 1s Sampling time Regarding the estimated extent, the errors for length, width
O"pos 0. 1 m· s- 2 Std. process noise
or 10 Time constant
and orientation follow the same trend as the Frobenius matrix
p 1 Scaling factor error does. FA is the worst algorithm, whereas MSULL
Np 10 3 Num. of particles and MSFFK have comparable average performance. Table II
(TP 1m Std. prediction axis min.
(T P
a Xmin summarizes the estimation errors (the best outcomes are
ax max 2m Std. prediction axis max.
(T P 2° Std. prediction orientation highlighted in boldface text). Figure 4 provides a histogram
a
representation of the Frobenius matrix errors for the four
updates, confirming the above analysis and ranking of the
about 0.005 s to process a single scan (negligible with respect updates.
to the particle filter-based computational burden depicted in
VI. CONCLUSIONS
Fig. 3(b». One can note that the kinematics (i.e. position and
velocity) are well estimated by all of the four updates. Very In this paper, a study of multi-sensor Bayesian extended
small differences can be seen for the position estimates before target tracking was presented. The extended targets were mod-
the convergence of the algorithms. However, these are limited eled in the random matrix framework. Four different extent
to less than 1 m. matrix measurement updates were proposed, and each uses

Authorized licensed use limited to: ULB Darmstadt. Downloaded on May 22,2025 at 20:16:33 UTC from IEEE Xplore. Restrictions apply.
Downloaded from [Link]
[Link] [Link]

the same multi-sensor kinematic vector measurement update. [17] c. Lundquist, K. Granstrom, and U. Orguner, "An extended target CPHD
One of the updates is based on a particle approximation of the filter and a Gamma Gaussian inverse Wishart implementation," IEEE
Journal of Selected Topics in Signal Processing, Special Issue on Multi-
extent state pdf, the other three are based on an inverse Wishart target Tracking, vol. 7, no. 3, pp. 472-483, Jun. 2013.
representation of the extent state pdf. Extensive numerical [18] R. Mahler, "PHD filters for nonstandard targets, I: Extended targets," in
results demonstrate that the best performance is obtained by Proc. of the 12th Intern. Con! on Inform. Fusion (FUSION) , Seattle,
WA, Jul. 2009, pp. 915- 921.
the PF-based approach paid by an increased computational [19] L. Mihaylova, A. Y. Carmi, F. Septier, A. Gning, S. K. Pang, and
burden. Comparable performance is observed for the two S. Godsill, "Overview of Bayesian sequential Monte Carlo methods for
updates based on multi-sensor generalization, and the worst group and extended object tracking," Digital Signal Processing, vol. 25,
pp. 1- 16, Feb. 2014.
performance is obtained by the updated based on fusion [20] U. Orguner, " A variational measurement update for extended target
approximation. tracking with random matrices," IEEE Trans. Signal Process., vol. 60,
In future research the theoretical differences and similarities no. 7, pp. 3827-3834, Jul. 2012.
[21] G. Vivone, P. Braca, K. Granstrom, A. Natale, and J. Chanussot,
between the four updates will be analyzed. Further simulation "Converted measurements random matrix approach to extended target
study will evaluate how the different updates scale with the tracking using X-band marine radar data," in Proc. of the 18 th Intern.
number of sensors and the magnitude of the sensor noise. Con! on Inform. Fusion (FUSION), Washington, US, Jul. 2015.
[22] G. Vivone, P. Braca, K. Granstrom, and P. Willett, "Multistatic bayesian
extended target tracking," Submitted to IEEE Trans. Aerosp. Electron.
REFERENCES Syst.,2015.
[23] M. Wieneke and J. W. Koch, "A PMHT approach for extended objects
[1] T. Ardeshiri, U. Orguner, and F. Gustafsson, "Bayesian inference and object groups," IEEE Trans. Aerosp. Electron. Syst. , vol. 48, no. 3,
via approximation of log-likelihood for priors in exponential pp. 2349- 2370, Jul. 2012.
family," CoRR, vol. abs!l510.01225, 2015. [Online]. Available:
[Link] 10.01225
[2] M. Arulampalam, S. MaskeU , N. Gordon, and T. Clapp, "A tutorial
on particle filters for online nonlinear/non-Gaussian Bayesian tracking,"
IEEE Trans. Signal Process. , vol. 50, no. 2, pp. 174-188, Feb. 2002.
[3] M. Baum and U. D. Hanebeck, "Extended object tracking with random
hypersurface models," IEEE Trans. Aerosp. Electron. Syst., vol. 50, no. 1,
pp. 149- 159, Jan. 2014.
[4] Y. Boers, H. Driessen, 1. Torstensson, M. Trieb, R. Karlsson, and
F. Gustafsson, "Track before detect algorithm for tracking extended
targets," lEE Proc. - Radar, Sonar Navig., vol. 153, no. 4, pp. 345-
351 , Aug. 2006.
[5] O. Cappe, S. GodsiU, and E. Moulines, "An overview of existing
methods and recent advances in sequential Monte Carlo," Proc. IEEE,
vol. 95, no. 5, pp. 899- 924, 2007.
[6] B. Errasti-Alcala and P. Braca, "Track before Detect algorithm for
tracking extended targets applied to real-world data of X-band marine
radar," in Proc. of the 17th Intern. Con! on Inform. Fusion (FUSION),
Salamanca, Spain, Jul. 2014.
[7] M. Feldmann, D. Franken, and J. W. Koch, "Tracking of extended
objects and group targets using random matrices," IEEE Trans. Signal
Process., vol. 59, no. 4, pp. 1409- 1420, Apr. 2011.
[8] K. Granstrom , A. Natale, P. Braca, G. Ludeno, and F. Serafino, "PHD
extended target tracking using an incoherent X-band radar: Preliminary
real-world experimental results," in Proc. of the 17th Intern. Con! on
Inform. Fusion (FUSION) , Salamanca, Spain, Jul. 2014.
[9] - -, "Gamma Gaussian inverse Wishart probability hypothesis density
for extended target tracking using X-band marine radar data," IEEE
Trans. Geosci. Remote Sens. , vol. 53, no. 12, pp. 6617- 6631 , Dec. 2015.
[10] K. Granstrom and U. Orguner, "A PHD filter for tracking multiple
extended targets using random matrices," IEEE Trans. Signal Process. ,
vol. 60, no. 11 , pp. 5657- 5671, Nov. 2012.
[11] K. Granstrom and U. Orguner, " A New Prediction Update for Ex -
tended Target Tracking with Random Matrices," IEEE Transactions on
Aerospace and Electronic Systems, vol. 50, no. 2, Apr. 2014.
[12] K. Granstrom and O. Orguner, "New prediction for extended targets with
random matrices," IEEE Trans. Aerosp. Electron. Syst., vol. 50, no. 2,
pp. 1577- 1589, Apr. 2014.
[13] A. K. Gupta and D. K. Nagar, Matrix variate distributions, ser. Chapman
& Hall/CRC monographs and surveys in pure and applied mathematics.
Chapman & Hall, 2000.
[14] J. W. Koch, "Bayesian approach to extended object and cluster tracking
using random matrices," IEEE Trans. Aerosp. Electron. Syst., vol. 44,
no. 3, pp. 1042- 1059, Apr. 2008.
[15] J. W. Koch and M. Feld mann, "Cluster tracking under kinematical
constraints using random matrices," Robot. Auton. Syst. , vol. 57, no. 3,
pp. 296- 309, Mar. 2009.
[16] J. Lan and X. R. Li, "Tracking of maneuvering non-ellipsoidal extended
object or target group using random matrix," IEEE Trans. Signal
Process., vol. 62, no. 9, pp. 2450- 2463, May 2014.

Authorized licensed use limited to: ULB Darmstadt. Downloaded on May 22,2025 at 20:16:33 UTC from IEEE Xplore. Restrictions apply.