Received: 23 October 2022

DOI: 10.1049/rsn2.12374

ORIGINAL RESEARCH
- -Revised: 11 December 2022 Accepted: 25 December 2022

- IET Radar, Sonar & Navigation

Multiple extended target tracking based on distributed

multi‐sensor fusion and shape estimation

Jinlong Yang1,2 | Mengfan Xu1,2 | Jianjun Liu1,2 | Fangdi Li1,2

1
School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, China
2
Jiangsu Provincial Engineering Laboratory of Pattern Recognition and Computational Intelligence, Jiangnan University, Wuxi, China

Correspondence Abstract
Jinlong Yang, School of Artificial Intelligence and Distributed multi‐sensor fusion based on the generalised covariance intersection (GCI)
Computer Science, Jiangnan University, Wuxi
214122, China.
fusion has been widely integrated into the Random Finite Set theory, which is promising
Email: [email protected] for multi‐target tracking with an unknown number of targets. However, it has not been
widely investigated in the multiple extend target tracking (METT) field, and there is still
Funding information an open problem on how to solve the inconsistency of label space among the sensors. For
Natural Science Foundation of Jiangsu Province, these problems, we first introduce the GCI fusion into the METT and proposed an
Grant/Award Number: BK20181340; National
Natural Science Foundation of China, Grant/Award
association algorithm by considering the estimated shapes and the target positions to
Number: 61305017 avoid the phenomenon of the label inconsistency as well as to reduce the computational
burden. Simulation results show that the proposed algorithm has a better tracking per-
formance than the traditional METT algorithms.

KEYWORDS
adaptive filters, multi‐target tracking, sensor fusion, signal processing

1 | INTRODUCTION extension) are usually unknowable. Mahler first proposed the

Random Finite Set (RFS) theory and extended the RFS‐based
With the development of modern sensor technology, the res- probability hypothesis density (PHD) filter to this field,
olution of sensor is gradually improving, which makes the forming the multiple extended target PHD algorithm (ET‐
original point target tracking become an extended target PHD) [6]. Then, Granström proposed the Gaussian mixture
tracking (ETT). The extended target is defined as that the version of ET‐PHD as the closed‐form solution, referred to as
target occupies multiple resolution cells of the sensor and/or ET‐GM‐PHD [7], and some improved versions are proposed
generates multiple measurements in one scan [1, 2]. A major in [8–10]. Compared to the traditional data association‐based
difficulty of the ETT is that the extent of target is unknown methods, such as Joint Probabilistic Data Association (JPDA)
and needs to be estimated in addition to the target position and algorithm [11], Multiple Hypothesis Tracking (MHT) algorithm
velocity. Recently, ETT has drawn considerable attention [3–5], [12], etc., these PHD‐based algorithms can approximate the
especially the multiple extended target (MET) tracking first order of the posterior distribution without considering the
(METT) [6–10]. The techniques of METT have been widely complex data association techniques.
used in the important military and civilian applications, For the shape estimation of METT, the random matrix
including missile defence system, aerial enemy reconnaissance (RM) was considered in [5], which has been used to track
and precision guidance, as well as traffic navigation and video elliptical target extension under the framework of ET‐GM‐
surveillance. PHD, forming the Gaussian Inverse Wishart PHD (GIW‐
For the METT with unknown number of targets, the PHD) filter [13–15]. Different from the RM‐based methods,
number of measurements and the shape information (physics the random hypersurface model (RHM) [16, 17] and star‐

-
This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is
properly cited.
© 2023 The Authors. IET Radar, Sonar & Navigation published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology.

IET Radar Sonar Navig. 2023;17:733–747. wileyonlinelibrary.com/journal/rsn2 733

734
- YANG ET AL.

convex RHM [3, 4, 18] were adopted to estimate the star‐ 2 | BACKGROUND
convex target extension and other general shapes. In [9, 19],
the B‐spline curve technique and Gaussian surface matrix‐ 2.1 | ET‐GM‐PHD
based methods were proposed to efficiently estimate the
irregular shapes of METs. Furthermore, the B‐spline curve was The ET‐GM‐PHD filter was established by employed
integrated into GIW‐PHD for non‐ellipsoidal‐shaped METT. Gaussian mixture (GM) approximation. Its prediction equa-
However, these algorithms only can be used for the scenarios tions of ET‐GM‐PHD filtering are referred to [6], and the
with known prior newborn density of targets and cannot be equation of updating is
used to classify the estimated METs, which limit their appli-
cation. Moreover, the shape features cannot be fully utilised for
state update to improve the tracking accuracy, especially for the Dkjk ðxÞ ¼ LZk ðxÞDkjk−1 ðxÞ ð1Þ
scenario with closely spaced METs. To address these problems,
we have proposed a Shape‐driven METT and a classification
where Dkjk ðxÞ represents the PHD‐intensity of state x, and
algorithm in previous work [20]. In this paper, we will intro-
LZk ðxÞ indicates pseudo‐likelihood function, and Zk represents
duce the association technique for identifying the newborn
measurement set at time k.
targets and reducing the computational burden on the basis of
Pseudo‐likelihood function is divided into two cases ac-
the previous work and extend it to the distributed multi‐sensor
cording to the measurement, when Zk ¼ ∅
fusion framework for further enhancing the tracking accuracy
of METT.
Recently, the RFS filtering based on the distributed multi‐ LZk ðxÞ ≜ 1 − pD ðxÞ þ e−rðxÞ pD ðxÞ ð2Þ
sensor (DMS) network is promising for multi‐target tracking
by using the generalised covariance intersection (GCI) fusion Otherwise,
proposed in [21], which can effectively improve the capabil-
ities of detection and tracking for moving targets, such as the � �
DMS‐based PHD algorithm [22–24], the DMS‐based Multi‐ LZk ðxÞ ≜ 1 − 1 − e−rðxÞ pD ðxÞ þ e−rðxÞ pD ðxÞ
Bernoulli (MB) and Labelled MB (LMB) algorithms [25].
ð3Þ
The GCI fusion is also called the exponential mixture density X X rðxÞjWj ϕ ðxÞ
(EMD), which is capable of using both Gaussian and non‐ ∗ ωP ∏z∈W z
P∠Zk W ∈P
dW λk c k
Gaussian multi‐target distributions. The DMS network also
can use arithmetic average (AA) fusion, which complies with
the law of ‘majority preference’ [26]. However, there is still an where ϕz ðxÞ represents the spatial distribution of the mea-
open problem on how to solve the label inconsistency phe- surements x, λk determines the number of clutter rate, and ck
nomenon among the sensors. Because the labels are inde- represents the spatial distribution of clutter. Assuming that the
pendently generated from the different sensors, the same clutter is Poisson‐distributed, λk ck represents the clutter den-
target in different sensors may be assigned the different sity. P ∠ Zk represents that partition P divides the measure-
labels. ment set Zk into cells W . Cell W contains more than one
Moreover, so far as we know, there is no DMS‐based RFS measurement that can bePviewed as coming from one extended
filtering for METT. Therefore, in this paper, we extend the target. The summation P∠Zk is taken over all partitions P of
P
ET‐PHD filter to DMS for METT and solve the label the measurement set Zk , the summation W ∈P is taken over
inconsistency problem by a proposed association method all cells W of partition P, and the product ∏z∈W is over all
integrating the estimated shape and position information. The measurements z in the cell W . rðxÞ represents the expected
main contributions include: (1) the irregular shapes of METs number of measurements and jW j represents the number of
are estimated by the B‐spline curve fitting technique; (2) an measurement points in W .
association method for matching the consistent targets be-
tween the two sensors by integrating the estimated shapes and ∏w∈p d W
ωP ¼ P ð4Þ
p∠Zk ∏w∈p d W
positions; (3) The GCI fusion is implemented for the distri-
butions of the consistent targets, forming the proposed
" #
algorithm.
ϕ ðxÞ
The rest of this article is organised as follows. Section 2 d W ¼ δjW j;1 þ Dkjk−1 e −γðxÞ
γðxÞ jW j
P D ðxÞ ∏ z ð5Þ
summarises ET‐GM‐PHD, B‐spline curve fitting technology Zk ∈W λk ck

and sensor fusion algorithms. Section 3 introduces the pro-

posed method, including the B‐spline shape estimation where d W represents the likelihood of partition W , ωP rep-
method, the association method based on shape estimation resents the weight of partition P. δjW j;1 represents a Kronecker
and distance information, and the track fusion method. The function.
performance of the proposed method is compared in Sec- When using distance partition in [7] and implementing
tion 4, and the conclusions are provided in Section 5. Gaussian mixing techniques to obtain a closed‐form solution, a
YANG ET AL.
- 735

standard ET‐GM‐PHD filter is obtained. However, this Establishing a pseudo‐measurement set Z ^ k is to move
method cannot estimate the target shape, and the distance the centre of the extended target to the coordinate origin
partition method cannot accurately divide the closely spaced without considering the concrete position of the extended
targets, and the number of partition schemes increase with the targets.
number of distance thresholds. Let Zk ¼ fzk;i gi¼1;2…;n as the measurement set, and jZk j
denotes the number of the measurements at time k, zk;i
denotes the ith measurement in zk . Define pseudo‐
measurement set Z ^ ¼ fz − xo g
2.2 | Irregular shape estimation based on k k;i i¼1;…;jZkþ1 j, xo denotes the
B‐spline fitting centre of extended target, Z ^ is the polar representation of
k
Zk .
In this section, a polar coordinate system is established with Update the extended target shape by pseudo‐measurement
the centroid of the extended target shape as the polar point. set Z^ .
k

8
>
> ðu − ui Þ2
>
>
> ; ui ≤ u < uiþ1
>
> ðuiþ1 − ui Þðuiþ2 − ui Þ
>
>
>
< ðu − ui Þðuiþ2 − uÞ þ ðu − uiþ1 Þðuiþ3 − uÞ ;
>
>
uiþ1 ≤ u < uiþ2
Bi;3 ðuÞ ¼ ðuiþ2 − uÞðuiþ2 − uiþ1 Þ ðuiþ2 − uiþ1 Þðuiþ3 − uiþ1 Þ ð8Þ
>
>
>
>
>
> ðuiþ3 − uÞ2
>
> ðu − u Þðu − u Þ; uiþ2 ≤ u < uiþ3
>
> iþ3 iþ1 iþ3 iþ2
>
>
:
0; other

The polar radius of the target shape is obtained by calculating 2.2.1 | Generate angle set fθi gni¼1 and the ith
the B‐spline shape of the extended target. B‐spline is a ^ .
partition measurement set Zk;i
smoothly segmented curve used to generate two‐dimensional
graphics. The B‐spline consists of a B‐spline basis function � n n uniform distributed polar angles in ½0; 2π� by angle set
Giving
and control vectors. θi gi¼1 . Then partition the� pseudo‐measurement set Z ^ as Z
^
k k;i
n
The B‐spline function is defined as a combination of according to the angle set θi gi¼1 , which can be expressed as
flexible bands by controlling vectors b, which can be expressed n � � � � � o
as Z^ ¼ Z ^ ��d Z ^ ; β < λ; κ Z ^ ; αi ¼ 1 ð9Þ
k;i k;j k;j i k;j
8
� � < 1; αi ⋅ ^z k;j > 0 and θi ∈ ð0; π�
n
X ^ ; αi ¼
κ Z k;j or αi ⋅ ^z k;j < 0 and θi ∈ ðπ; 2π� ð10Þ
bðuÞ ¼ pi Bi;E ðuÞ; u ∈ ½uE−1 ; uEþ1 � ð6Þ :
i¼0
−1; others
� �2
� �
where pi in P k ¼ ½p1 ; p2 ; ⋯; pn �T is a control point and Bi;E ðuÞ � �αi ⋅ ^z k;j �
d ^z k ; αi ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð11Þ
is the E − 1 degree of the B‐spline basis function, u is defined ðαi ÞT ⋅ αi
in the knot vector U ¼ fu1 ; u2 ; ⋯; uE ; ⋯; unþE g. Bi;E ðuÞ can
be generated recursively by where βi ¼ ½−tanðθi Þ; 1� and αi ¼ ½1; tanðθi Þ�. βi is a line along
the ith partition angle θi and
� through
� the origin of coordinate,
� while αi is vertical to βi . d ^z k;j ; βi denotes the distance from
1; ui < u < uiþ1
Bi;1 ðuÞ ¼ ^zk;j to the line βi , and λ is the width of partition. k ⋅ k2 rep-
0; others �
resents 2‐norm. Moreover, κ ^zk;j ; αi is a constraint condition
u − ui uiþE − u
Bi;E ðuÞ ¼ Bi;E−1 ðuÞ þ Biþ1;E−1 ðuÞ used to confirm the set Z ^ .
uiþE−1 − ui uiþE − uiþ1 k;i

ð7Þ
2.2.2 | Generate polar radius set fρi gni¼1
2 X � �
When E ¼ 3, the B‐spline basis function is expressed by ρi ¼ �� �� d Z^ ; αi
^
^zk;j ∈Zk;i k;j ð12Þ
Equation (8). ^ �
�Z k;i
The detailed process of irregular shape estimation based on
B‐spline fitting is given in [19]. where j ⋅ j denotes the number of elements.
736
- YANG ET AL.

2.2.3 | Convert the representation of the set of

control points
Transform the expression of polar radius set fρi gni¼1 as P k
from the polar coordinate system to cartesian coordinate sys-
tem and generate a smooth curve related to the shape of the
measurement set by applying the B‐spline function of Equa-
tion (6) through these new control points set P k .
Figure 1 shows the measurement shape by the B‐spline
fitting. Black dots indicate the pseudo‐measurement set Z ^ ,
k
^
the partition Zk;i is constituted by the area enclosed with points
O1, O2, O3, and O4. The curves indicate the B‐spline shape. βi
is a line along the ith partition angle θi and through the coor-
FIGURE 1 Measurement shape by the B‐spline fitting.
dinate origin, while αi is vertical to βi . λ is the width of partition.

where w1 and w2 denote the relative fusion weights for each

2.3 | GCI fusion posterior satisfying w1 þ w2 ¼ 1. The Metropolis weights are
defined and given by Equation (11) in [24], which can be used
The GCI concept for distributed multi‐sensor fusion is pro- for the information averaging among neighbour sensors. The
posed in [21], which has been widely extended to the RFS GCI fusion rule in the form of Equation (14) can be derived by
theory for distributed multi‐target tracking. In this section, we minimising the weighted sum of Kullback‐Leibler divergence
will describe the concept of the GCI fusion and its drawbacks. (KLD) [23],

πw ¼ argminðw1 DKL ðπkπ1 Þ þ w2 DKL ðπkπ2 ÞÞ ð15Þ

π
2.3.1 | GCI fusion
where DKL is the Kullback‐Leibler divergence defined as
The GCI fusion amounts to computing the density that min-
imises the sum of the information gained from local posteriors, Z
f ðXÞ
avoiding the problem of double‐counting of the common in- DKL ðf kgÞ ¼ f ðXÞlog dX ð16Þ
formation. Assume that at time k, there are two local multi‐ gðXÞ
target posterior probability densities conditioned on the mea-
surements
� � �obtained �from� Sensors
� 1 and 2, denoted by R f ðXÞ and gðXÞ are different probability distributions,
where
k � 1:k k � 1:k
and ⋅dX is the set integral.
π1 X �Z1 and π2 X �Z2 . Then, the fused posterior
� � �
�
probability density πw Xk �Z1:k ; Z1:k
is defined as the fusion Remark In the real scenario, although the local filter works
1 2
well, the GCI fusion may fail to obtain the accurate results due
of the two posterior probability densities, that is, to the drawback of label inconsistencies. The label inconsis-
tency implies that the same target may be assigned to different
� � � � � �
� k � 1:k labels in different sensors, which is usually caused by the
πw Xk �Z1:k
1 ; Z 1:k
2 ¼ π 1:k
w X �Z1 ∪ Z2 ð13Þ
following reasons, the measurement‐driven birth process for
the different sensors, the uncertainty observations with the
It is noted that if the two sensors have the same noise and clutter interference, etc. Therefore, we should fuse
� �observation
�
� the distributions of the consistent targets between the two
noise for the same targets, then the π1 Xk �Z1:k 1 and
� � � sensors. In this paper, we will propose an association method
�
π2 Xk �Z1:k2 have certain correlations with the common noise for identifying the consistent targets between two sensors by
comprehensively considering the shape and position infor-
distribution. If we directly fuse them without considering their
mation of the estimated targets. Once they are right associated
unknown correlations, the problems about the correlations
and labelled, then the labels of the survival targets between the
would arise. In [21], Mahler proposed the GCI fusion rule,
neighbour sensors can be easily identified later, because their
which extends the Finite‐Set Statistics (FISST) to the distributed
trajectories can be maintained by the filtering method in each
environments. Under this extension, the fused posterior is the
sensor.
geometric mean, or the exponential mixture of the local pos-
teriors, that is,
� �
k � 1:k
� w1 � �
�
�w 2 2.3.2 | GCI fusion with PHD
� �
�
� π 1 X �Z 1 π2 Xk �Z1:k 2
πw Xk �Z1:k 1:k
1 ; Z2 ¼ R � �� �w 1 � � � w2 ð14Þ
k � 1:k Sensor ID is denoted by s, s ∈ S. Assuming Sensor 1 and 2 are
π1 Xk �Z1:k
1 π 2 X �Z 2 δX
neighbours, the targets in the Sensor 1 can be fused with the
YANG ET AL.
- 737

targets associated in the neighbouring Sensor 2, and the 3 | PROPOSED MULTI‐SENSOR

associated targets are considered as survival targets whose la- FUSION ALGORITHM FOR METT
bels are consistent, while targets that are not associated may be
newborn targets or clutters. The process of GCI fusion is as The GCI fusion mentioned above has the problems of label
follows. inconsistency and computational inefficiency; especially, the
Assume Ls ¼ fℓis gi¼1:n denotes the track label set of the label inconsistency will lead to a significant reduction in
targets in sensor s, and L ¼ fLs gs¼1:Ns is the set of track labels tracking accuracy. To address these problems, we propose an
of all targets in all the sensor, n denotes the number of the optimal distributed multi‐sensor MTT algorithm under the
track label type, and Ns denotes the number of sensors. Then, GM‐PHD framework with two main processes.
the label expression for the posterior density of the targets in Process 1: Target tracking and trajectories maintenance in
sensor s can be represented as each sensor. Before density fusion of the targets between the
� � � neighbouring sensors, the estimated components in each sensor
� � �
ðℓis Þk ��Zðs ℓs Þ1:k ; s ¼ 1; …; N
i
� and latest measurements need to be performed the data asso-
πs Xk �Z1:k
s ¼ ⋃ π s X � s
i¼1:n ciation based on position and shape information to distinguish
ð17Þ the survival targets, the newborn targets and/or clutters. Espe-
cially, the tracks of the survival targets are maintained in each
sensor.
Assume that there are some labels in sensor 1 associated by Process 2: Posterior density fusion between the neigh-
� � 2, we define the mapping function as τð⋅Þ,
the labels in sensor bouring sensors. The densities of survival targets between the
j
that is, ℓi1 ¼ τ ℓ2 denotes that the target with label ℓi1 in neighbouring sensors can be directly fused according to the
j consistent labels. The consistent labels can be directly identi-
sensor 1 is associated by the target with ℓ2 in sensor 2, which
fied according to the trajectory maintenance. For the densities
means that they have the label consistency. The targets in the
of the newborn targets, they will be fused by using the Murty
label matching set are considered as the survival target, and the
algorithm [27]. It is noted that the Murty algorithm is only
GCI fusion based on the label matching is given by
executed once for each newborn target during the initial stage

� � � � � i �
� � ðℓ1 Þ ðℓj2 Þ
πw X�ZL1 1 ; ZL2 2 ¼ ⋃ �
πw X�Z1 ; Z2
i¼1:jL1 j

j¼1:jL2 j
j
ℓi1 ¼τðℓ2 Þ

� � � � � j �w2 ð18Þ
� i w1
ℓi1 Þ � ðℓ1 Þ
j �
ℓ2 Þ � ðℓ2 Þ
ð
π1 X �Z1 ð
π2 X �Z2
¼ ⋃ � � j �w1 � � j �w2
R j � j �
i¼1:jL1 j ð ℓ1 Þ � ð ℓ1 Þ ð ℓ2 Þ � ð ℓ2 Þ
π1 X �Z1 π2 X �Z2 δX
j¼1:jL2 j
j
ℓi1 ¼τðℓ2 Þ

where j ⋅ j denotes the number of the track label type. of the trajectory. After the trajectory initiation, the newborn
For the targets that they appear at the newborn time and/ targets will become the survival targets and their trajectories
or they are clutters, if we fuse them according to the GCI, the will be maintained in each sensor.
fuse performance will decrease. Here, we will distinguish them
at the next time by using the association method, if they can be
associated with the targets in the neighbour sensor, they will be 3.1 | B‐spline shape estimation
considered as the real newborn targets and allocated the
matching labels to join the fusion. It is noted that the distri- 3.1.1 | Initialise shape set X
butions of these targets without fusion at the newborn time
will be fed back to the filters for avoiding miss tracking of the Assume X ^ represents a B‐spline shape, X^ ¼ fρ ; θi g20 , here ρ
i i¼1 i
next time. and θi represent the ith pole radius and polar angle, respectively.
738
- YANG ET AL.

The
� 1 initial � angles are ∅ ¼ fθ1 ; …; θi ; …; θ20 g ¼
polar Assume that P is a set of peaks, θp is a set of angle loca-
10 π; 2
10 π; …; 2π , and their corresponding polar radii can be tions of peak P, B is the set of bottoms, θB is the set angle
set as the same length, for example, ρ1 ¼ ρ2 ¼ … ¼ ρ20 . locations of bottom B, and j ⋅ j is the number of set elements.
� � � � ��
θp ; P ¼ θp1 ; P 1 ; ⋯; θpi ; P i ; ⋯; θjPj ; P jPj
3.1.2 | Update shape set X n � � �o ð20Þ
ðθB ; BÞ ¼ ðθB1 ; B1 Þ; ⋯; θBj ; Bj ; ⋯; θjBj ; BjBj
Updating X by calculating the B‐spline shape (Equation 6). � �
where bottom point θBj ; Bj is the neighbour of the peak
�
3.1.3 | Maximum shape estimation likelihood point θpi ; P i . If the distance between peak P i and the
ϕX ð⋅Þ neighbour bottom Bj is small and their polar angle difference is
small, then we define this peak as pseudo‐peak with label
For better estimation of the similarity of shapes, the 20 polar Pseui ¼ 1, that is,
radii can be extended to 360 polar radii, for example, � � � �
i¼1 → fρgi¼1 . The likelihood of shapes X1 and X2 can be
fρg20 Pseui ¼ 1; �P i − Bj � < P th and θPi − θBj < θth
360
ð21Þ
calculated by
� �
� n1 − n2 �
�� where P th and θth are the thresholds
� (e.g.,�P th ¼ 0:3,
� θth ¼ 50).
ϕX ðX1 ∣ X2 Þ ¼ exp −�
� � When Pseui ¼ 1, remove θpi ; P i and θBj ; Bj from point
maxðn1 ; n2 Þ� sets of peaks and bottoms.
ð19Þ
X
360 � � �2 �
� exp −�ρi2 − ρi1 �
i¼1 3.1.5 | New maximum shape estimation
likelihood ϕX ð⋅Þ
where n1 and n2 represent the number of peaks when the
shapes X1 and X2 are stretched out according to the polar The polar radius, polar angle and number of peaks should be
angles. Figure 2 shows the stretch‐out view of one extended made full use to obtain a more accurate shape estimation
target shape. function ϕX ð⋅Þ.
� � ��
� n1 − n2 �
ϕX ðX1 ∣ X2 Þ ¼ exp −�
� � �
3.1.4 | Find the peak location of the stretch‐out maxðn1 ; n2 Þ�
view of the extended target shape minX
ðn1 ;n2 Þ
ð22Þ
� � �2 � � ��
exp −�ρi2 − ρi1 � � exp −�θi2 − θi1 �
Some pseudo‐peaks can affect the accuracy calculated by i¼1
Equation (19). Figures 2 and 3 show the correct peaks by
removing the Pseudo‐peaks according to Equation (21). The
where n1 and n2 denote the number of peaks in X1 and X2 ,
curves descript the 360° expansion of the B‐spline shape,
respectively. The shape similarity essentially depends on the
where one pseudo‐peak exists in the solid curve and is
corresponding peaks, which requires a matching strategy to
removed forming the dotted curve with correct peaks.

FIGURE 2 Remove Pseudo‐peak of case 1. FIGURE 3 Remove Pseudo‐peak of case 2.

YANG ET AL.
- 739

match the peaks of the two shapes. The cycle shifting method where μ is a constant parameter depending on the speed of
is selected as the matching strategy, which is referred to [20]. the target, the subscripts Z and J of the components denote
they are from Zk and Jk , respectively. miZ denotes the position
of Zk .
j
3.2 | Association method based on shape It is noted that for targets XiZ and XJ , if their measure-
and distance ments
� are close
� to each other (e.g., the position weight
j
For the METT, the shape information is extremely important dist miZ ; mJ > 0:5) and far from the other targets (e.g., their
as the identification feature for discriminating the different q jmJ j
position weights distðmiZ ; mJ Þq¼1;q≠j < 0:5), there is no need to
target with different shapes, especially for the closely spaced
METT. In light of these advantages, we propose an improved consider the shape information. Therefore, we do not need to
ET‐GM‐PHD filter by integrating the estimated shape infor- calculate ϕX ð⋅Þ and set ϕX ð⋅Þ ¼ 1, which can effectively save
mation; especially, we propose an association method for the computational burden. If there are some other targets close
j
identifying the diffident targets, such as the survival targets, to the targets XiZ and XJ , shape estimation should be consid-
newborn targets, and the clutters, which can be conducive to ered, which can be calculated by Equation (25).

8
> � � � �jmJ j
� < 1 dist miZ ; mjJ > 0:5 and dist miZ ; mqJ
� > < 0:5
j
ϕX X iZ ∣ X J ¼ � � � � � �
q¼1;q≠j ð25Þ
>
: NX X iZ ∣ X jJ P X X iZ ∣ X jJ ∅X X iZ ∣ X jJ
> otherwise

measurement update, reducing the computational burden and Where

better identifying the closely spaced targets. Moreover, it also 0 � �1
can adaptively identify newborn targets without the fixed birth � � � i j �
i j � nZ − n J �
model for the newborn targets in the traditional ET‐GM‐PHD NX X Z ∣ X J ¼ exp −�
@ � � � �A
�
filter. �max niZ ; njJ �

j
�
3.2.1 | Calculate weights � � min niZ ;nJ � � �2 �
j
X � i;n j;n �
� � P X X iZ ∣ XJ ¼ exp −�ρZ − ρJ �
j
The weight w Zik ; J k defined as Equation (23) denotes the n¼1
j
similarity of position and shape between Zik and J k . Zk denotes j
�
� niZ ;nJ
� minX � � ��
measurement set, Jk denotes predicted target set, and the su- j � j;n �
perscripts i and j denote the ith and jth components in the ∅X X iZ ∣ X J ¼ exp −�θi;n
Z − θJ �
n¼1
respective sets, that is, Zik denotes the ith components in the
measurement subset Zk . The state positions are denoted as � �
j
m ¼ ðx; yÞ and the B spline shapes are denoted as NX X iZ ∣ X J denotes the likelihood of the number of
� �
X ¼ fρi ; θi gni¼1 . j
peaks, P X X iZ ∣ X J denotes the likelihood of the peak height,
� �
j
�
j
� �
j
� �
j
� and ∅X X iZ ∣ X J denotes the likelihood of the peak polar
w Zik ; J k ¼ dist miZ ; mJ � ϕX X iZ ∣ X J ð23Þ
angles. The superscript n denotes the nth components of the
set, that is, ρi;n i
Z denotes the nth peak height in ρZ .
0 rffi�ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�
ffiffiffiffiffiffiffiffiffiffiffi�ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�
ffiffiffiffi 1
i j 2 i j 2
� � B xZ − xJ þ yZ − yJ C
j
dist miZ ; mJ ¼ expB
@ − C
A 3.2.2 | Filtering algorithm based on association
μ
A two‐dimensional association matrix A is established by
ð24Þ calculating the weight að⋅Þ between the measurement subset Zk
740
- YANG ET AL.

and the predicted component set Jk . Here, the calculation as the clutter measurement, and we need remove them from
method is the same as wð⋅Þ of Equation (23). the measurements set, that is,

2 � �3 n o � �
a Z1k ; J 1k … a Z1k ; J vk j j
6 � � 7 Z0b;k ¼ Zb;k − Zk ; count Zk < c ð29Þ
j
A¼6
4… a Zik ; J k … 7
5 ð26Þ
� �
a Zuk ; J 1k … a Zuk ; J vk
Confirm newborn targets n o
At time k, b possibleðkÞ ¼ Z0b;k , placing the measurements
where u denotes the number of measurement subsets and v set into the suspected newborn target set birth possibleðkÞ.
denotes the number of prediction targets. At time k þ 1, associate the targets between Z0b;kþ1 and
� �
b possibleðkÞ by weight w Z0b;kþ1 ; b possibleðkÞ obtained
3.2.3 | Extract the survival targets according to Equation (23). If the value of some weights more
than the threshold (e.g., 0.5), then we consider that the targets
Following the principle that one predicted Gaussian compo- in birth possibleðkÞ corresponding to these weights are
nent should be associated with only one measure subset newborn targets.
(generated from one extended target), based on the association
matrix, the measurement with the highest association weight is
selected to update 3.2.5 | Assign track label to newborn target
� the �corresponding Gaussian component.
j
The elements a Zik ; J k of the association matrix in Equa-
For a newborn target, the label v can be assigned for recording
tion (26) can be updated by the track, and the state of the target at time k can be expressed
as
8 � � � � ��
j j
>
>
> 1; if a Zik ; J k ¼ max a Zrk ; J k
� � < r¼1:u ζk ¼ fx k ; Xk ; Vk g ð30Þ
j � �
a Zik ; J k ¼ a Z i j ð27Þ
>
> and k k ≥ wth
; J n ojJk j
>
: j
0; otherwise where the label set Vk = vk denotes Gaussian compo-
j¼1
nent labels set at time k, jJk j denotes the number of predicted
� � � � �� � �
j j j components at time k. x k and Xk denote state and shape of the
where a Zik ; J k ¼ max a Zrk ; J k denotes that a Zik ; J k
r¼1:u Gaussian components, respectively, at time k.
has the maximum value for all the measurement set corre-
sponding to the jth Gaussian component. wth denotes the
threshold of weight (e.g., wth ¼ 0:5). 3.3 | Distributed fusion of the consistent
extended targets
3.2.4 | Extract the newborn targets For traditional multi‐sensor fusion methods based on GCI,
generally the position information of the target is only
Establish newborn targets set Zb;k considered, while the important extended information of the
According to Equation (27), if some measurements cannot target is not used. Moreover, the Gaussian components with
match the predicted targets, that is, these measurements cor- low weights are used for fusion, the fusion accuracy may be
responding to zero weight in associated matrix A maybe affected and the calculational cost is high.
newborn targets or clutters. We will add these measurements to To alleviate these problems, the shape estimation infor-
the subset of candidate newborn target measurements Zb;k . mation is considered comprehensively. We propose an opti-
mising distributed multi‐sensor MTT algorithm, which
� � � �v �
� i j employs an association algorithm that combined the estimated
Zb;k ¼ Zb;k ∪ Zik �w
Zk ; J k ¼0 ð28Þ shapes and the target position to avoid the phenomenon of the
j¼1
label inconsistency as well as to reduce the computational
burden. Through this association method, the survival targets
where v ¼ jJk j denotes the number of the components from will be associated between the two neighbour sensors, and their
the predicted set Jk . trajectories have been maintained in each sensor. In other
words, if they can be matched (right associated), they can be
Remove the clutters from the newborn targets set fused directly by GCI. Otherwise, they are newborn targets or
countð⋅Þ
� �is the number of measurement points, if missed estimation in one sensor; we cannot directly fuse them
j j
count Zk < c, c is a threshold (e.g., 3), then Zk is considered avoiding the wrong fusion. In this proposed algorithm, for
YANG ET AL.
- 741

these survival targets without being right associated, we directly Calculate the matched label set of the newborn targets
consider them as the part result of the fusion and directly between neighbour sensors
feedback for the next filter. Moreover, for the newborn targets, It is noted that all the targets are the newborn targets at the
the Murty's method [27] is employed to search the match newborn time; after this label matching, their labels can be
relation of them between the neighbour sensor, and the GCI is identified according to the trajectory maintenance of the sur-
implemented for the matched targets like the fusion process vival targets. At the next time, only the newborn targets need to
for the survival targets. continue to make the label matching. If sensor 1 has p newborn
The steps of the proposed algorithm are as follows: we targets recorded in X1 and sensor 2 has q newborn targets
continue to use s ∈ S as the sensor ID, and si denotes the ith recorded in X2 , calculate their association matrix A similar to
sensor. Equation (26) and weight að⋅Þ similar to Equation (27), that is,
L ¼ fLs gs¼1:Ns is the finite label space of the GM com-
ponents from all the sensor, Ls ¼ f ℓis gi¼1:n denotes the 2 � q� 3
a X 11 ; X 12 … a X 11 ; X 2
track label set of the targets in sensor s, and ℓis indicates the 6 �
j
� 7
track label i of targets from the sensor s. n denotes the number A¼6
4… a X i1 ; X 2 … 7
5 ð33Þ
of types of the track label, and Ns denotes the number of p � p q�
a X1; X2 1
… a X1; X2
sensors.

aðX1 ; X2 Þ ¼ distðm1 ; m2 Þ � ϕX ðX1 ∣ X2 Þ ð34Þ

3.3.1 | METT for each sensor

Assume that Sensor 1 and Sensor 2 are neighbour sensors, and The Murty's method in [27] is employed to derive the
minimum cost association matrix A, which can obtain the
the posterior density of the multi‐targets
� � s is
in� sensor
matched label set.
i � ðℓis Þ
expressed by Equation (17), and there πs Xð s Þ ��Zs
ℓ
can be
Get the mismatched label set between the individual
expressed by
neighbour sensor
After minimising cost association matrix A in (2.2), there may
� � � � i � be labels that mismatched the labels between the neighbouring
i � ð ℓi Þ ðℓi Þ ðℓ Þ ð ℓi Þ ð ℓi Þ
πs Xðℓs Þ ��Zs s ¼ ωs s ∗ N xs s ; ms s ; Ps s lef t
sensors s1 and s2 , which are denoted as L1 and L2 .
lef t

� i � ð31Þ
ðℓ Þ ðℓi Þ
∗ ϕ Xs s ∣ Zs s
3.3.3 | Fuse the posterior distribution πw and
ðℓi Þ
πleft
s of the consistent targets and inconsistent
ωs s denotes the weight of predicted target with label ℓis , distribution posterior
Nð⋅Þ denotes the Gaussian likelihood of position of target
with label ℓis , and ϕð⋅Þ denotes the likelihood of shape of Fuse the posterior distribution πw of the consistent targets
target with label ℓis . jℓs j denotes the number of the label types The targets involved in the fusion are matching survival targets
j
of sensor s. with consistent labels ℓi1 and ℓ2 , sensor 1 is viewed as a local
sensor (fusion centre) and sensor 2 is viewed as a neighbour
sensor. Therefore, they can be fused directly by the GCI, thus
3.3.2 | Identify the consistent targets between obtaining an optimal fused solution.
two sensors The posterior probability of Sensor 1 can be expressed by
� � � X � �
Define label match and label mismatch ðL Þ � ðL Þ J1 i ðℓi Þ ðℓi Þ ðℓi Þ
j π1 X1 1 �Z1 1 ¼ i¼1 ω1 N x1 1 ; m1 1 ; P1 1
L1 ð1; 2Þ ¼ f ℓi1 j ℓi1 ¼τðℓ2 Þgi∈L1 ;j∈L2 denotes the label set that
j � i � i �
all the labels ℓi1 in sensor 1 can be matched with the labels ℓ2 in ðℓ Þ � ðℓ Þ
neighbour sensor 2. L1 denotes the labels out of matched
left ∗ ϕX X1 1 ��Z1 1
label set in sensor 1, which is mismatched with any label from
ð35Þ
sensor 2.
Then, the labels of the sensor 1 can be expressed as follows
�Then,
� GCI fusion
� of Equation (14) can be expressed as
� ðL1 Þ L2
L1 ¼ L1 ð1; 2Þ ∪ Lleft ð32Þ πw X�Z1 ; Z2 , that is,
1
742
- YANG ET AL.

� � � � � i �
� ðL Þ ðL Þ � ð ℓ Þ ð ℓj Þ
πw X�Z1 1 ; Z2 2 ¼ ⋃ πw X��Z1 1 ; Z2 2
i¼1:J1

j¼1:J2
j
ℓi1 ¼τðℓ2 Þ

� � i �w1 � � j �w2
ℓi1 Þ � ðℓ1 Þ ℓ2 Þ � ð ℓ2 Þ
j
ð
π1 X �Z1 π2 X ð �Z2
¼ � � i w1 � � � j �w2
R ð ℓi1 Þ � ðℓ1 Þ ð ℓ2 Þ � ð ℓ2 Þ
j
π1 X �Z1 π2 X �Z2 δX
�� i � � i � i ��w1 � � j � � j � j ��w2
ðℓ1 Þ ðℓi1 Þ ðℓi1 Þ ð ℓ1 Þ � ð ℓ1 Þ j ðℓ2 Þ ðℓj2 Þ ðℓj2 Þ ðℓ Þ � ðℓ Þ
ωi1 N
x1 ; m1 ; P1 ϕX X1 �Z1 ω2 N x2 ; m2 ; P2 ϕX X2 2 �Z2 2
¼ � � i � � i � i ��w1 � � j � � j � j ��w2
R i ðℓ1 Þ ðℓi1 Þ ðℓi1 Þ ð ℓ1 Þ � ð ℓ1 Þ j ðℓ2 Þ ðℓj2 Þ ðℓj2 Þ ðℓ Þ � ðℓ Þ
ω1 N x1 ; m1 ; P1 ϕX X1 �Z1 ω2 N x2 ; m2 ; P2 ϕX X2 2 �Z2 2 δX

ð36Þ

The detail of Equation (36) is in Appendix. J1 and J2 denote " #−1

ððℓi1 Þ;ðℓj2 ÞÞ ðℓi1 Þ −1 ðℓj2 Þ
−1
the number of label types of sensor 1 and sensor 2, respectively. Pð1;2Þ ¼ w1 P1 þ w2 P2 ð40Þ
Equations (37) and (38) are simplified formulas used for
the calculation of GCI fusion πw ð⋅Þ, that is,
!w
X X
πi ¼ πwi ð37Þ "
i i ððℓi1 Þ;ðℓj2 ÞÞ ððℓi1 Þ;ðℓj2 ÞÞ ðℓi1 Þ −1 ðℓi1 Þ
mð1;2Þ ¼ Pð1;2Þ w1 P1 m1
ω1 N ðx; m1 ; P1 ÞϕX ðX1 jZ1�Þω2 N ðx; m� 2 ; P2 Þϕ
� X ðX2 jZ2 Þ # ð41Þ
ð38Þ
≈ ωð1;2Þ N x; mð1;2Þ ; Pð1;2Þ ϕX Xð1;2Þ �Zð1;2Þ ðℓj2 Þ
−1
ðℓj2 Þ
þ w2 P2 m2
Then, a Gaussian representation of the CGI fusion for-
mula is obtained by Equation (39).

� � � � ��
ððℓi1 Þ;ðℓj2 ÞÞ ððℓi1 Þ;ðℓj2 ÞÞ ððℓi1 Þ;ðℓj2 ÞÞ ððℓi1 Þ;ðℓj2 ÞÞ �� ððℓi1 Þ;ðℓj2 ÞÞ
� � i
� ðℓ1 Þ ðℓj2 Þ
� ωð1;2Þ N x; mð1;2Þ ; Pð1;2Þ ϕX Xð1;2Þ �Zð1;2Þ
�
πw X�Z1 ; Z2 ¼ � � � � �� ð39Þ
R ððℓi1 Þ;ðℓj2 ÞÞ ððℓi Þ;ðℓj ÞÞ ððℓi Þ;ðℓj ÞÞ ððℓi Þ;ðℓj ÞÞ � ððℓi Þ;ðℓj ÞÞ
ωð1;2Þ N x; mð1;2Þ1 2 ; Pð1;2Þ1 2 ϕX Xð1;2Þ1 2 ��Zð1;2Þ1 2 dx

� � � j �w2 � � � �
The subscript indicates the sensor ID and the superscript ððℓi Þ;ðℓj ÞÞ ðℓi1 Þ w1 ðℓ Þ ðℓi1 Þ ðℓj Þ
ωð1;2Þ1 2 ¼ ω1 ω2 2 ϵ w1 ; P1 ϵ w2 ; P2 2
indicates the pair of matching targets in both sensors, for
� �
ððℓi Þ;ðℓj ÞÞ ðℓi Þ ðℓj Þ ðℓi Þ ðℓj Þ
example, ωð1;2Þ1 2 denotes the weight of pair of matching ∗ N ω1 1 − ω2 2 ; 0; P1 1 w1 −1 þ P2 2 w2 −1
j
targets with label ℓi1 from sensor 1 and label ℓ2 from sensor 2. ð42Þ
YANG ET AL.
- 743

qffiffiffiffiffiffiffiffiffiffiffiffiffiffi−1ffiffiffiffi
detð2πPw Þ
where ϵðw; PÞ ¼ ½detð2πPÞ�w . association method in the proposed algorithm. Then, it is
� � � continue to fuse the position and shape estimations between the
� neighbour sensors by the proposed association method.
πleft
s X ðℓleft Þ � ðℓleft Þ
Z
� s : ℓleft ∈ L left
The parameters of the motion and measurement models
0 � 1
� are given by the following equations.
m � ð ℓm Þ
¼ ⋃ πleft @Xðℓleft Þ �Zs left A
s �
m¼1:jℓleft j � xkþ1 ¼ Fxk þ Gwk
ð43Þ ð44Þ
� � zk ¼ Hxk þ vk
ð ℓm Þ ðℓm Þ ðℓm Þ ðℓm Þ
¼ ⋃ ωs left ∗ N x s left ; ms left ; P s left 2 3
m¼1:jℓleft j 1 0 T 0
� m � 60 1 0 T7
ðℓleft Þ ðℓmleft Þ F ¼6
40
7 ð45Þ
∗ ϕ Xs ∣ Zs 0 1 0 5
0 0 0 1
� �
1 0 0 0
Posterior distribution πleft
of the newborn targets and the
s
H¼
0 1 0 0
ð46Þ
missed estimation targets
Let jℓ1 j represents the number of track labels types in Sensor 1
where F and H are the transition matrix and the measurement
Nm represents the number of track labels types in Sensor 1
matrix, respectively. x k and z k are the state vector and the
matched with the tracks in Sensor 2. If 0 ≤ Nm < jℓ1 j or
observation vector at time k, respectively. wk and vk are the
0 ≤ Nm < jℓ2 j, it indicates that some targets in sensor 1 are not
process noise and observation noise, respectively.
matched with those in its neighbour sensor. These targets may
The sensor sampling period is T ¼ 1 s and the target
be newborn targets or missed estimation targets by the
moves in a straight line at a constant speed. The probability of
neighbour sensor. At this time, for the Nm matched labels, the
detection P D ¼ 0:99 and the probability of survival P s ¼ 0:99.
same fusion algorithm as that in case (3.1) is performed to
For each step, the Poisson rate of the generated target mea-
obtain the fused distribution πw .
surement is γ ¼ 10. Measurement noise covariance is
For the labels left, they cannot be fused directly, the
R k ¼ diag½0:3; 0:3� and the process noise covariance is
Murty's method is employed for searching the match relation
Q ¼ diag ½1; 1�. Two different metrics were used to evaluate
of them between the neighbour sensor, and the GCI is
the algorithm performance. The first one is the Optimal Sub‐
implemented for the matched targets like the fusion process
patten Assignment (OSPA) distance [28] and the second one is
for the survival targets. If some targets are not matched, the
the operation time.
distribution of them will be directly feedback to the filters for
The birth process is modelled as a Poisson random finite
the next filtering avoiding to missing estimation.
set with Gaussian mixing intensity.
These inconsistent targets include labels that cannot be
matched in Sensor 1 Lleft 1 and Sensor 2 L2 . Here, we call all
left
D b ðx ; X Þ¼
the inconsistent target labels as L . Then, their posterior
left
� n � �
1 ð47Þ
distribution Equation (31) in their individual sensor is shown in ω1b N x ∣ m1b ; P 1b þ σ b I � ∏ N X b ∣ Xb;i ; ∇1b;i
ð ℓm Þ i¼1
Equation (43). There, ωs left denotes the weight of predicteds
target with label ℓmleft , Nð⋅Þ denotes the Gaussian likelihood of where m1b ¼ ½−500 m; 20 m; 200 m=s; 2 m=s�T , P 1b ¼ diag½100;
position of target with label ℓm left , and ϕð⋅Þ �denotes � the likeli-
1 1
400; 100; 400�, ω1b ¼ 0:1; n ¼ 20, X b;1 ¼ ⋯ ¼ X b;n ¼ 10,
hood of shape of target with label ℓm left . �ℓleft � denotes the ∇1b;1 ¼ … ¼ ∇1b;n ¼ 1.
number of labels from ℓleft . It is noted that the newborn targets in the following two
scenes are adaptive identified for the proposed algorithm,
while for the other comparison algorithms, ttheir newborn
4 | SIMULATIONS intensities are given in advance, that is, they are prior
knowledge.
To verify the effectiveness of the proposed algorithm, the GCI‐
GM‐PHD [24] and AA‐GM‐PHD [26] algorithms were selected
as the comparison algorithms. These two algorithms only can 4.1 | Linear motion scene with 3 targets
provide the fused position estimation without shape estimation,
and they cannot adaptive estimate the newborn targets. While In this experiment, assume that there are three targets making a
the proposed algorithm has the abilities of irregular shape esti- cross motion in a two‐dimensional scenario of
mation, newborn target identification and label consistency as- ½−500; 500� � ½0; 300� and last 50 s. The motion parameters
sociation, the ET‐GM‐PHD algorithm [7] is first implemented and the motion trajectories of the targets are shown in Table 1
in each sensor by integrating the proposed B spline‐based shape and Figure 4. The experimental results are shown in
estimation method and the shape‐based measurement Figures 5–8 and in Table 3.
744
- YANG ET AL.

As can be seen from Figure 5, the proposed algorithm has It is noted that the 3rd second, the proposed algorithm and
the least OSPA error compared with the other algorithms from the ET‐GM‐PHD based on association algorithm have similar
the 10th second to the 50th second, which means that the high OSPA errors. The reason is that two newborn targets
proposed algorithm has the highest tracking accuracy. Espe- appear at the 3rd second affecting the tracking accuracy,
cially, for the intersection point of time 42, the targets 2 and 3 because the position of newborn targets is not given in
can be well distinguished and tracked with the help of shape advance to the above two algorithms, while the other algo-
information, and thus, there is no sharp peak of OSPA for the rithms have the position of newborn targets.
proposed algorithm. The ET‐GM‐PHD algorithm based on In Figure 6, the proposed algorithm has high accuracy for
shape information and data association in each sensor also has cardinality estimation almost all the time. However, at the 3rd
a high tracking accuracy between times 42 and 50. By com- second, the proposed algorithm has low accuracy for cardi-
parison, the accuracies of GCI‐GM and AA‐GM algorithms nality estimation because of the same reason above.
are lower. The reason is that the targets at the intersection As can be seen from Figure 7, for the proposed algorithm,
point cannot be well distinguished and tracked, and one target the shapes of targets 1, 2 and 3 can be well estimated and the
disappears at 45th second without being identified, which af- two intersection targets 2 and 3 are well separated at 42nd
fects the tracking accuracy. second due to the proposed shape‐based association method.

TABLE 1 The motion parameters of the

Target Initial position (m) Time of birth Time of death Speed (m/s) Time of crossing
3 targets
1 (−500,200) 1 45 (20,2) None

2 (500,200) 3 50 (−12,−2) 42

3 (−470,288) 3 50 (10,−4) 42

FIGURE 4 Motion trajectories of the extended targets. FIGURE 6 Cardinality estimation.

F I G U R E 5 The Optimal Sub‐pattern Assignment (OSPA) statistics

for different algorithms. FIGURE 7 Filtering results of the proposed method.
YANG ET AL.
- 745

However, the AA‐based algorithm without using the shape‐ As can be seen from Figure 10, the proposed algorithm has
based association method cannot separate the intersection the least OSPA error compared with the other algorithms from
targets well at time 42 shown in Figure 8. Therefore, it has a the 10th second to the 50th second, which means that the pro-
lower tracking accuracy than the proposed algorithm. posed algorithm has the highest tracking accuracy, followed by
In Table 3, it is clear that although tracking accuracy has the ET‐GM‐PHD based on the association algorithm. The GCI‐
been greatly improved, its time consumption is not increased, GM and AA‐GM algorithms have lower tracking accuracy. The
similar to those of the other algorithms. The reason is that the reason is that the proposed shape‐based data association method
label match operator only needs be performed once in the is implemented in the proposed algorithm, but not in the GCI‐
proposed algorithm for the newborn targets at newborn time GM and AA‐GM algorithms. The targets crossed at 19th, 35th,
to keep the label consistence between the sensors, while the and 42nd seconds and some targets disappearing at 30th and
matching relation of the survival targets has been recorded 45th seconds also affect the tracking accuracy.
according to the label history of trajectory maintenance
without needing the label match operator, which can greatly
reduce the computational cost of the matching computation of
the proposed algorithm. Moreover, the use of the proposed
data association method also reduces the time consumption of
the proposed algorithm, especially the time for measurement
update, because the predicted targets do not need to be
updated by all the measurements, but only need to be updated
by the associated measurements.

4.2 | Linear motion scene with 5 targets

Assume there are five targets making a cross motion in a two‐

dimensional scenario and last 50 s. The motion parameters and
the motion trajectories of the targets are shown in Table 2 and FIGURE 9 Motion trajectories of the extended targets.
Figure 9. The experimental results are shown in Figures 9–13
and in Table 3.

F I G U R E 8 Filtering results of the arithmetic average‐Gaussian F I G U R E 1 0 The Optimal Sub‐pattern Assignment (OSPA) statistics
mixture‐probability hypothesis density method. for different algorithms.

TABLE 2 The motion parameters of the

Time
5 targets
of
Target Initial position (m) birth Time of death Speed (m/s) Time of crossing
1 (−500,200) 1 45 (20,2) None

2 (500,200) 3 50 (−12,−2) 19,35,42

3 (−470,288) 3 50 (10,−4) 42

4 (100,0) 5 50 (0,4) 35

5 (200,100) 10 30 (10,6) 19
746
- YANG ET AL.

TABLE 3 Estimation of time consumption between different

algorithms

GM‐PHD + Proposed
Algorithms AA (s) GCI (s) connection(s) algorithm (s)
3 targets 2.5622 2.5628 2.5764 2.6331

5 targets 3.7189 3.6472 3.2619 3.4548

reduce the computational cost of the matching computation of

the proposed algorithm.
It is noted that the 3rd, 5th, and 10th seconds, the pro-
posed algorithm and ET‐GM‐PHD based on the association
algorithm have similar high OSPA errors. The reason is that
two newborn targets appear at the 3rd, 5th, and 10th seconds
FIGURE 11 Cardinality estimation. affect the tracking accuracy due to the same reason above.
Moreover, as can be seen from Figure 11, the proposed
algorithm has high accuracy for cardinality estimation almost
all the time. At the 3rd, 5th, and 10th seconds, the proposed
algorithm has low accuracy for cardinality estimation duo to
the same reason above.
At 19th, 35th, and 42nd seconds, GCI‐GM and AA‐GM
without the shape‐based association algorithm cannot find
the corresponding targets to track because of the interference
of the crossed targets, which leads to the low weights of pos-
terior probability and low accuracy for cardinality estimation.
As can be seen from Figure 12, for the proposed algo-
rithm, the intersection points of targets 2, 3, 4 and 5 can be
correctly separated, and their shapes can be estimated well at
19th, 35th, and 42nd seconds, while the AA‐based algorithm
without using the association method cannot separate the
crossed targets well shown in Figure 13.

FIGURE 12 Filtering results of the proposed method.

5 | CONCLUSION

In this paper, we propose a novel multiple extended target

tracking algorithm based on the distributed multi‐sensor
fusion. First, the B spline‐based shape estimation method is
employed for irregular shape estimation, and then the esti-
mated shape and position information are considered for
measurement association in each sensor and label consistency
association between the neighbour sensors. Thus, we can
adaptive estimate the newborn targets in each sensor and
match the consistency labels between the neighbour sensors.
Especially, it is more effective in avoiding closely spaced tra-
jectory labelling errors. Finally, the estimated targets in each
sensors are fused by using the CGI method if they have con-
sistency labels. Simulations show that the proposed algorithm
has a good performance in irregular shape estimation of METs,
F I G U R E 1 3 Filtering results of the arithmetic average‐Gaussian
mixture‐probability hypothesis density method. trajectory fusion tracking, and reduction of GCI fusion time.

In the proposed algorithm, the label match operator only A UT HO R C O N TR I B U T I O N S

needs to be performed for the newborn targets to keep the Jinlong Yang and Jianjun Liu initiated the project and designed
label consistence between the sensors, while the matching the algorithms. Mengfan Xu and Fangdi Li participated in the
relation of the survival targets has been recorded according to proposed method and performed the experiments. All authors
the label history of trajectory maintenance, which can greatly drafted and approved the final manuscript.
YANG ET AL.
- 747

ACKN OW LE DG E ME N T S 13. Granstrm, K., Bramst, J.: Bayesian smoothing for the extended object
This work is partially supported by the Natural Science random matrix model. IEEE Trans. Signal Process. 14(67), 3732–3742
Foundation of Jiangsu Province (No. BK20181340) and the (2019)
14. Tuncer, B., Ozkan, E.: Random matrix based extended target tracking
National Natural Science Foundation of China (No. with orientation: a new model and inference. IEEE Trans. Signal Process.
61305017). 69(69), 1910–1923 (2021)
15. Li, P., et al.: Modified Gaussian inverse Wishart PHD filter for tracking
CON FL ICT OF I N T ER E ST multiple non‐ellipsoidal extended targets. Signal Process. 150(9), 191–203
The authors declare that they have no conflicts of interest. (2018)
16. Gao, L., et al.: Bayesian approach to multiple extended targets tracking
with random hypersurface models. Radar Sonar & Navigation Iet 13(4),
DATA AVA IL AB I LI T Y S TAT E M EN T 601–611 (2019)
The data is available in the experiments of this article, which 17. Sun, L., et al.: Modeling and tracking of maneuvering extended object
can be generated according to the presented scenarios of the with random hypersurface. IEEE Sensor. J. 18(21), 20552–20562 (2021)
18. Sun, L., et al.: Maneuvering extended object tracking with modified star‐
experiments.
convex random hypersurface model based on minimum cosine distance.
Rem. Sens. 14(17), 4376 (2022)
O RC ID 19. Yang, J., Li, P., Ge, H.: Extended target shape estimation by fitting B‐
Jinlong Yang https://orcid.org/0000-0001-9548-4236 spline curve. J. Appl. Math. 2014, 1–9 (2014)
20. Li, F., Yang, J.: Shape‐driven multiple extended target tracking and clas-
REFE R ENC ES sification based on B‐Spline and PHD filter. In: International Conference
on Signal Processing Systems, Shanghai (2022)
1. Gilholm, K., Salmond, D.: Spatial distribution model for tracking 21. Mahler, R.: Optimal/robust distributed data fusion: a unified approach.
extended objects. In: IEEE Proceedings‐Radar, Sonar and Navigation, In: Proceedings of SPIE 4052, Signal Processing, Sensor Fusion, and
vol. 152, no. (5), pp. 364–371 (2005)
Target Recognition IX, pp. 128–138, Orlando (2000)
2. Gilholm, K., Godsill, S., Salmond, D.: Poisson models for extended
22. Li, G., et al.: Distributed multi‐sensor multi‐view fusion based on
target and group tracking. In: Proceedings of the Signal and Data Pro- generalized covariance intersection. Signal Process. 166, 107246 (2020)
cessing of Small Targets, pp. 230–241. SPIE, CA, USA (2005) 23. Leonard, M.R., Zoubir, A.M.: Multi‐target tracking in distributed sensor
3. Ma, T., et al.: Tracking of maneuvering star‐convex extended target using
networks using particle PHD filters. Signal Process. 159(159), 130–146
modified adaptive extended Kalman Filter. IEEE Access 8(8), (2019)
214030–214038 (2020) 24. Li, T., Corchado, J.M., Sun, S.: On generalized covariance intersection for
4. Wang, L., et al.: Joint tracking and classification of multiple extended distributed PHD filtering and a simple but better alternative. In: 20th
targets via the PHD filter and star‐convex RHM. Digit. Signal Process.
International Conference on Information Fusion, pp. 1–8. IEEE, Xi'an,
111, 102961 (2021) China (2017)
5. Koch, J.W.: Bayesian approach to extended object and cluster tracking 25. Li, S., et al.: Multi‐sensor multi‐object tracking with different fields‐of‐
using random matrices. IEEE Trans. Aero. Electron. Syst. 44(3), view using the LMB filter. In: Proceedings of 21st International Con-
1042–1059 (2008)
ference on Information Fusion (FUSION), pp. 1201–1208. IEEE,
6. Mahler, R.: PHD filters for nonstandard targets, I: extended targets. In: Cambridge (2018)
Proceedings of the 12th International Conference on Information 26. Li, T., Corchado, J., Sun, S.: Partial consensus and conservative fusion of
Fusion, pp. 915–921. IEEE, Seattle (2009)
Gaussian mixtures for distributed PHD fusion. IEEE Trans. Aerosp.
7. Granstrom, K., Lundquist, C., Orguner, U.: Extended target tracking Electron. Syst. 55(5), 2150–2163 (2019)
using a Gaussian‐mixture PHD filter. IEEE Trans. Aero. Electron. Syst. 27. Murty, K.G.: An algorithm for ranking all the assignments in order of
48(4), 3268–3286 (2012) increasing cost. Oper. Res. 16(3), 682–687 (1968)
8. Granström, K., et al.: Likelihood‐based data association for extended
28. Schuhmacher, D., Vo, B.T., Vo, B.N.: A consistent metric for perfor-
object tracking using sampling methods. IEEE trans. intell. veh. 3(1), mance evaluation of multi‐object filters. IEEE Trans. Signal Process.
30–45 (2018) 56(8), 3447–3457 (2008)
9. Yang, J., et al.: Multiple extended target tracking algorithm based on
Gaussian surface matrix. J. Syst. Eng. Electron. 27(2), 279–289 (2016)
10. Yang, J., et al.: An improved ET‐GM‐PHD filter for multiple closely‐
spaced extended target tracking. Int. J. Control Autom. Syst. 15(1),
468–472 (2017) How to cite this article: Yang, J., et al.: Multiple
11. Fortmann, T., Bar‐Shalom, Y., Scheffe, M.: Sonar tracking of multiple extended target tracking based on distributed multi‐
targets using joint probabilistic data association. IEEE J. Ocean. Eng. sensor fusion and shape estimation. IET Radar Sonar
8(3), 173–184 (1983)
Navig. 17(5), 733–747 (2023). https://doi.org/10.1049/
12. Blackman, S., Dempster, R.J., Broida, T.J.: Multiple hypothesis track
confirmation for infrared surveillance systems. IEEE Trans. Aero. rsn2.12374
Electron. Syst. 29(3), 810–824 (1993)