A Hybrid Maneuvering Extended Target Tracking Algorithm Based on PMBM Filter

d
Xingxiang Xiea , Yang Wanga,∗, Junqi Guoa , Rundong Zhoub

we
a Harbin Institute of Technology Shenzhen, Shenzhen, 518055, China
b Harbin Institute of Technology, Harbin, 150001, China

ie
Abstract
This paper presents a hybrid maneuvering extended target tracking algorithm based on the Poisson multi-Bernoulli
mixture (PMBM) filter in the random finite sets (RFS) framework. The proposed method consists of forward propagation

ev
of the PMBM filtering density, followed by backward propagation of a multi-Bernoulli (MB) filtering density. In this
paper, we first provide a standard PMBM forward recursive, and then give a detailed derivation of the MB backward
smoothing. Moreover, a closed-form solution to the proposed hybrid algorithm is given via an implementation of Gamma
Gaussian inverse-Wishart (GGIW) technology. In addition, the Murty’s algorithm is employed to relieve the complexity
in this work. The performance of the proposed method is evaluated in a simulation study.

r
Keywords: Poisson multi-Bernoulli mixture, forward-backward smoother, Gamma Gaussian inverse-Wishart, random
finite sets, extended target tracking. er
1. Introduction probability hypothesis density (PHD) filter [12], the cardi-
nalized PHD (CPHD) filter [13], the cardinality balanced
Multi-target tracking (MTT) refers to the process of multi-Bernoulli (CB-MeMBer) filter [14], and the MTT
pe
jointly estimating the number of targets, states and their conjugate priors, e.g., labelled multi-Bernoulli (LMB) fil-
trajectories based on noisy sensor data [1]. MTT has been ter [15], the generalized labelled multi-Bernoulli (GLMB)
widely used in many different scenarios, e.g., autonomous filter [16], the PMBM filter [17], [18]. General RFS fil-
driving [2], pedestrian tracking [3], multi-robot simultane- ters consist of prediction achieved by using Chapman-
ous localization and mapping (SLAM) [4], mono-camera Kolmogorov equation [19] and filtering achieved by us-
3D multi-object tracking [5], X-band marine radar systems ing Bayesian formula. This work considers the problem
[6].
ot

of smoothing based on the RFS framework in ETT.

Generally, MTT consists of extended target tracking Prediction, filtering and smoothing are three key in-
(ETT) which means that each target generates multiple terrelated steps of a dynamic system for jointly targets
measurements per time step, i.e., multiple resolution cells detection and tracking conditioned on noisy sensor mea-
are occupied by a target and point target tracking (PTT)
tn

surements. Prediction provides a means of forecasting the

which means that each target generates only a single mea- state of the system at some point in the future conditioned
surement per time step, i.e., a single resolution cell is oc- on measurements collected up to the present time by the
cupied by a target [7]. The focus of this paper lies on operator, whereas filtering aims at estimating the state of
ETT. Indeed, some approaches to ETT are built on the the system at regular time steps based on all the mea-
base of complex data association problems such as typ- surements collected up to that instant [20]. Smoothing
rin

ical joint probabilistic data association (JPDA) [8] and typically refers to estimating the target state of the past
multiple hypothesis tracking (MHT) [9]. The RFS the- time using all measurements till the current time.
ory [10], a novel and rigorous approach has been devel- In other words, they amount to calculating
oped by Mahler based on the finite set statistics (FISST),
represents a powerful tool to cope with these problems fk|l (xk |z1:l ) (1)
ep

mentioned above. More detailed descriptions of the RFS

concepts and of FISST are given in the books [10], [11]. the probability density of the state xk at time k given
For decades, a variety of RFS filters based on multi-target the observation history z1:l = (z1 , . . . , zl ) up to time l.
Bayesian filter framework have been proposed, including Prediction, filtering and smoothing, respectively, refer to
the cases l < k, l = k, and l > k [21],[22].
The Bernoulli smoother modeled the target in one of
Pr

∗ Corresponding author two “present” or “absent” modes by a Bernoulli RFS

Email addresses: [email protected]
(Xingxiang Xie), [email protected] (Yang Wang), is first proposed in [23], moreover, also gives a deriva-
[email protected] (Junqi Guo), [email protected] tion of multi-Bernoulli smoother by an approximation to
(Rundong Zhou)

Preprint submitted to Elsevier December 25, 2022

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
 the multi-target smoothing recursion using multi-Bernoulli
Table 1: Notation

d
RFS. In [24], a novel Bernoulli forward-backward smooth-
ing algorithm incorporated the track-before-detect (TBD) No. Symbols Explanation
sensor model with multi-scan information is presented. 1 x single extended target state

we
For multi-target tracking problem, some smoothing 2 X extended targets state set
algorithms have been proposed in the literature, e.g., 3 y smoothed extended targets state
the forward–backward PHD smoothers are proposed in 4 Y smoothed extended targets state set
[25, 26, 27, 28], but the PHD filters are known to suf- 5 z single extended target measurement
fer from highly uncertain cardinality estimates since they 6 Z extended targets measurement set
model the state of targets as a Poisson RFS, and the same 7 f p (·) PPP RFS density
8 f b (·)

ie
as the PHD smoothers. The implementation of a tractable Bernoulli RFS density
approximation of the usual CPHD smoother is presented, 9 f mb (·) MB RFS density
which can provide stable estimates in the number of tar- 10 f mbm (·) MBM RFS density
gets, but bears a higher computational cost [29]. How- 11 f pmbm (·) PMBM RFS density

ev
ever, all of these smoothers mentioned above use unla- 12 GAM (·) Gamma distribution
belled RFS to model the multi-target state, thus they can- 13 N (·) Gaussian distribution
not be used to directly estimate target trajectories. For 14 IW (·) inverse-Wishart distribution
directly estimating smoothed target trajectories, a gen- 15 > Matrix Transpose
eral multi-trajectory forward-backward smoothing equa-

r
tion is first derived in [30] based on sets of trajectories and
the RFS framework. The closed-form solution to multi- n 
o
target forward-backward smoothing recursion based on the
er the parameters: λu , Wj, rj,i , pj,i i∈Ij
. In the
j∈J
GLMB filter is proposed in [31], but shows highly uncer- process of backward propagation, the backward smoothing
tain of tracking when new born targets are an independent density is given by computing in reverse order from o k−1
and identically distributed (IID) cluster RFS or Poisson
n

to 0 with the MB filtering density: rj,i , pj,i i∈Ij .
RFS [32]. In [33], a forward–backward LMB smoother is j∈J
pe
presented, which has distinct advantages for maintaining 2) The closed-form solution of the presented method is
the independence of different trajectories and their trajec- provided via the GGIW implementation. We also propose
tory outputs, but cannot solve the problem of trajectory a more efficient alternative to deal with the data associ-
fragmentation [34] when the label of a track changes before ation problem based on Murty’s algorithm, which is con-
the track ends. structed by the existence probability of each component.
In [35], a one time step lagged δ-GLMB smoother 3) A strategy of complexity reduction is adopted to
is proposed, which is capable of estimating the number effectively reduce computational cost of the hybrid al-
ot

of targets and outputs of targets trajectories, but bears gorithm, which refers to recycling Bernoulli components
the super-exponentially growing number of filtering com- which are less than a fixed threshold and approximates
ponents and numerical complexity. A forward-backward their distribution as PPP, which are incorporated into
smoother based on the LMB RFS for the evaluation of the PPP intensity represented undetected targets in the
tn

autonomous driving systems is proposed in [36], but can- PMBM RFS framework.
not guarantee enough information for the initialization of The remainder of this paper is as follows. Section
newborn targets. Furthermore, more smoothing solutions II summarizes relevant concepts in multi-target Bayesian
based on labelled RFS filters have been developed [34], filtering, forward–backward smoother and the RFS mod-
[37], [38]. elling. Section III gives the detailed derivation of the pro-
rin

Two extended target PMBM trackers are proposed in posed hybrid algorithm. Section IV introduces the GGIW
[39], but have not been implemented via smoothing-while- implementation of the presented method. Simulation re-
filtering GGIW technology in a non-linear model. To ad- sults are presented in Section V and conclusions are given
dress the problems mentioned above, we present a hy- in Section VI.
brid algorithm based on PMBM filter in the RFS frame-
2. BACKGROUND
ep

work. Lastly, we evaluate the performance of the proposed

method in a non-linear numerical simulation. The pro-
posed hybrid algorithm consists of two parts where one is In this section, we provide relevant background re-
the forward PMBM filtering recursion composed of pre- quired for the derivation of the main results, including
diction and update (i.e., filtering, one of the above three multi-target Bayesian filter, multi-target Bayesian for-
steps), and the other is backward MB smoothing recursion. ward–backward smoother and the RFS modeling. Nota-
Pr

This paper mainly contains the following contributions: tion is given in Table 1.
1) The detailed implementation of the proposed so-
lution is given. In the process of forward propagation,
we utilize the PMBM filtering recursion which propagate
2

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
 2.1. Multi-target Bayesian filter and forward–backward

d
smoothing recursion
The recursion of generic multi-target Bayesian for-
ward–backward smoother is shown in Fig. 1, which con-

we
sists of forward Bayesian filtering and backward Bayesian
smoothing. In this section, we give an overview as follows.
RFS is a recursive method based on Bayesian filter [10].
The extended target states and measurements are both
modeled as RFS. At each time step k, let x ∈ Rn denotes
a single extended target state, each target state x is also

ie
a random variable. The measurements set obtained at
Figure 1: The recursion of generic multi-target Bayesian for-
time step k is denoted as Zk = {z1 , . . . , zn }, including ward–backward smoother.
clutter and target generated measurements. The number
of targets, i.e., target set cardinality |Xk |, is a time-varying

ev
state with a transition density g (·|X) according to a
discrete random variable. Markov chain, or dies with probability 1 − pS (X). The
Let fk|k (Xk |Zk ), fk|k−1 (Xk |Xk−1 ) and fk (Zk |Xk ) de- multi-target set Xk are the union of the surviving targets
note the multi-target set density, the multi-target tran- and new targets, which are born independently following
sition density and the multi-target measurement like- a Poisson point process (PPP) with intensity λb (·).
lihood, respectively. The multi-target Bayesian filter

r
The multi-target state at time step k can be denoted as
propagates in time the multi-target posterior density Y = Y + ∪Y − , where Y + denotes the newborn targets and
fk−1|k−1 (Xk−1 |Zk−1 ) using the Chapman-Kolmogorov Y − denotes the surviving targets. They are disjoint and
prediction equation

fk|k−1 (Xk |Zk−1 )

Z
er independent. From Lemma1([33], Sec.2.2), the joint multi-
target Markov transition density fk|k−1 (·|X) is written as
[11]
fk|k−1 (Y |X) = fB,k|k−1 Y + fS,k|k−1 Y − |X .



= fk|k−1 (Xk |Xk−1 ) fk−1|k−1 (Xk−1 |Zk−1 ) δXk−1 (2) (6)
pe
Considering the survival targets only, multi-target Markov
and the Bayesian update formula transition density fS,k|k−1 is as follows
fS,k|k−1 Y − |X =


fk (Zk |Xk ) fk|k−1 (Xk |Zk−1 )
fk|k (Xk |Zk ) = R . (3)
fk (Zk |Xk ) fk|k−1 (Xk |Zk−1 ) δXk
X X Y pj,i −
S,k|k−1 fk|k−1 (Y |X)
1 − pS,k|k−1 . (7)
The generic multi-target Bayesian forward–backward ]Y j =X i i,j
1 − pj,i
S,k|k−1
ot

smoother (smoothing recursion) is described as follows

2.2.2. Extended target measurement model
fk−1|l (Xk |Zk ) = Let Zk = Zc ] Z1 ] · · · ] Zn is the measurement set,
fk|l (Y |Zk ) fk|k−1 (Y |Xk−1 ) where Zc , Z1 , · · · , Zn are mutual independent disjoint sets.
Z
tn

fk−1 (Xk−1 |Zl ) δY. (4) Zc modelled as a PPP with intensity λC (·) is the clut-
fl|l−1 (Y |Zl )
ter measurement set. z ∈ Zi is the measurement gener-
ated by target i (i = 1, · · · , n). Each extended target state
R
The sets of integral f (X) δX is defined as follows
[11], xk ∈ Xk is either detected with probability pD (xk ) or
missed with probability 1 − pD (xk ). The extended target
∞
rin

measurements are modelled as a PPP rate γ (xk ) and spa-

Z Z
X 1
f (X) δX = f ({x1 , · · · , xn }) dx1:n , (5) tial distribution ϕ (·|xk ), independent of any other targets
n=0
n!
and their corresponding generated measurements.
where f (·) is a real-valued function on the space Rn , if We use convolution formula for multiple extended tar-
f (·) > 0, its set integral is one. For simplicity the following get likelihood density, also known as the conditional ex-
tended target measurement set likelihood,
ep

inner product is adopted throughout the paper: Y

Z l (z|xk ) = pD (xk ) e−γ(xk ) γ (xk ) ψ (z|xk ). (8)
hf, gi = f (x) g (x) dx. z∈Z

For an extended target xk , the effective detection

2.2. Standard multi-target system models probability is the product of the detection probability
Pr

2.2.1. Multi-target dynamic model and the PPP probability density function
(PDF), i.e.,
pD (xk ) 1 − e−γ(xk ) , where 1 − e−γ(xk ) is the Pois-


As described in Sec.II A, for the extended target set
Xk = {x1 , . . . , xn } at time step k, each target x ∈ Xk ⊆ son probability of generating at least one measurement.
Rn survives with probability pS (X) and moves to a new Therefore, the effective probability that the extended tar-

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
 get xk is not detected is where f j,i (·) are Bernoulli RFS densities, see Sec.II C2.
Xkd is the set of detected targets. W j is the weight prob-

d
qD (xk ) = 1 − pD (xk ) + pD (xk ) e−γ(xk ) . (9) ability. The MBM RFS density is the normalized and
weighted sum of n multi-target densities of oMBs, which is

we
Note that the conditional likelihood (8) for an empty mea- j
 j,i j,i

parameterized by W , r ,p i∈Ij
, i.e., MBM
surement set, l∅ (xk ), is also described by (9) [40]. n o j∈J
RFS ∼


= W j , rj,i , pj,i i∈Ij .
2.3. RFS modelling j∈J

2.3.1. Poisson point process RFS 2.3.4. PMBM RFS

Two basic forms of RFS are the PPP and the Bernoulli The PMBM filter developed in recent years is an RFS

ie
process. A PPP is an RFS whose cardinality is Poisson method, which is defined as a convolution of a PPP inten-
distribution and for any finite cardinality, each element sity for the undetected targets and an MBM density for the
in the set is IID. The PPP intensity λ (x) = µf (x) is detected targets. The PMBM filter is an ETT conjugate
determined by the scalar Poisson rate µ and the spatial prior based on standard RFS method [41]. The signifi-

ev
distribution f (x). The PPP density is given by cance of conjugate prior is, given the assumed models, we
Y can know what the theoretically exact density is and find
f p (X) = e−µ µf (x) (10) computationally tractable approximations [42].
x∈X
In the PMBM model, the set of targets Xk is a union of
two disjoint set, i.e., Xk = Xku ]Xkd . The set of undetected

r
2.3.2. Bernoulli process RFS
targets denoted Xku , modelled by PPP. The set of detected
A Bernoulli process with probability of existence r and
targets denoted Xkd , modelled by MBM. The PMBM set
existence-conditioned PDF p (x) has RFS density
density can be represented by [40]

b

1 − r, X = ∅

er f pmbm (Xk ) =
X
f p (Xku ) f mbm Xkd ,


(15)
f (X) = rp(x) , X = {x} , (11)
 Xku ]Xkd =Xk
0, otherwise

u
pe
Y
f p (Xku ) = e−hλ ;1i
λu (x), (16)
where r ∈ [0, 1] is an existence probability, p (x) is a PDF x∈Xku

of target x. It can be noticed that one Bernoulli can be
X j X Y

f mbm Xkd = f j,i Xki .


W (17)
completely characterized by parameters r and p (x) , i.e., j∈J
U
X i =X d i∈Ij
i∈Ij
Bernoulli RFS ∼ = (r, p (x)).
Therefore, the PMBMn RFS density is defined o entirely
2.3.3. MBM process RFS
ot

u j
 j,i j,i

by the parameters λ , W , r , p i∈Ij
, i.e.,
A multi-Bernoulli (MB) RFS consists of multiple inde-  n o 
j∈J
pendent Bernoulli component, PMBM RFS ∼


= λu , W j , rj,i , pj,i i∈Ij .
[ j∈J
X= X i. (12)
tn

3. A HYBRID MANEUVERING EXTENDED

Furthermore, let I is the index set of the Bernoulli com- TARGET TRACKING ALGORITHM
ponents. J is the index set of the MBs. In the MBM, the The proposed hybrid algorithm consists of forward fil-
number of MB components is |J|, the jth MB component tering achieved by the PMBM filter recursion, followed by
has |Ij | Bernoulli components. The MB density is backward smoothing achieved by an MB density given in
rin

X Y the Sec.II C3, as shown in Fig. 2. In this section, we first

f mb Xkd = f j,i Xki .



(13) give a representation of PMBM forward filtering, then give
X i =X d i∈Ij
U
i∈Ij a derivation of an MB backward smoothing recursion.
Where ] denotes the disjoint union, the MB RFS den-
3.1. PMBM forward filtering
ep

sity refers to multiple Bernoullio RFS, which is pa-

The PMBM filter density is parametrized by the PPP
n
, i.e., MB RFS ∼


rameterized by rj,i , pj,i (x) i∈Ij =
n o j∈J intensity and the MBM density. Predicting and updating
the PMBM density means to compute the prediction and


rj,i , pj,i (x) i∈Ij . The MBM density is
j∈J update parameters [41]. In this section, based on the above

X j X Y PMBM RFS model and the set of measurements Zk , the
Pr

f mbm Xkd = f j,i Xki ,

W (14) PMBM forward filtering density recursion can be given as
X i =X d i∈Ij
U
j∈J i∈Ij follows.
Prediction Step: Given a posterior PMBM filtering
density at time k − 1 with parameters

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
 d
ie we
Figure 2: The proposed hybrid algorithm

ev
(   )
λuk (x) = qD (x) λuk|k−1 (x) ,
 
fk−1 ∼
= λuk−1 , j
Wk−1 , j,i
rk−1 , pj,i
k−1 , (18) (25)
i∈Ij j∈J
j Q
j
Wk|k−1 C∈A LC
the predicted
( PMBM filtering density is as follows WA,k =P P j Q , (26)
A∈Aj Wk|k−1 LC
  )
  j∈Jk|k−1 C∈A

r
∼
fk|k−1 = λ u
, W j
, r j,i
,pj,i
,
k|k−1 k|k−1 k|k−1 k|k−1
i∈Ij j∈J where Aj is the space of all data association hypothesis.
er (19) C is a subset of all data association A. A data association
A ∈ Aj is an assignment of each measurement in Zk to
where
a source, either to the background (clutter or new target)
λuk|k−1 (x) = λbk−1 (x) + λuk−1 (x) ; pS fk|k−1 (·|·) , (20a) or to one of the existing targets indexed by Ij . qD (x) is
the probability of missed detection. LC
is the likelihood
D E
j,i
rk|k−1 = pj,i (x) ; pS
j,i
rk−1 , (20b)
pe
D
k−1
E function of data association C . fCj X C are a Bernoulli
pj,i
k−1 (x) ; pS,k fk|k−1 (·|·)
density. For more details, please refer to [41].
j,i
pk|k−1 (x) = D E , (20c) Furthermore,
( (21) is described with parameters)as fol-
pj,i
k−1 (x) ; pS,k
 n o 
lows fk ∼= λuk , WA,kj
, rkj,i , pj,i
k , i.e.,
j j i∈Ij
Wk|k−1 = Wk−1 . (20d) j∈J
fk ∼
= f pmbm (Xk |Zk ).
ot

λb (x) is the intensity function of target birth, pS is 3.2. MB backward recursion

the survival probability and fk|k−1 (·|·) is the transition
For backward smoothing recursion, a generic approach
density.
to compute the smoothed density fk−1|l (·) is via (4). How-
Update Step: Given a prior PMBM filtering density
ever, calculating the set integral exactly in (4) is generally
tn

with parameters
intractable. Here we assumed that it does not take into
fk|k−1 ∼
= account the undetected targets (i.e., missed detection of
(  )
PPP). In order to obtain an appropriate solution to this
 n o
u j j,i j,i
λk|k−1 , Wk|k−1 , rk|k−1 , pk|k−1 j
, smoothing recursion, we use an MB density to compute
i∈I j∈J
backward smoothed filter fk−1|l (·) at time k = l − d (d
(21)
rin

is the fixedtime lag, generally,

 d = 1). We rewrite MB
 
a set of measurements Zk and the standard measure- j,i j,i
density as rk , pk j
. The backward smooth-
ment model [32], the updated PMBM filtering density at i∈Ik j∈Jk
time k is as follows ing recursion can be obtained as stated in the following
proposition.
X
f pmbm (Xk |Zk ) = f p (Xku ) f mbm Xkd , (22)


ep

Proposition: Given a posterior MB filter density at

Xku ]Xkd =Xk
time k − 1 with parameters
where   
(X) ∼
u j,i
, pj,i
Y mb
f p (Xku ) = e−hλk (x);1i λuk (x), (23) fk−1 = rk−1 k−1 j
, (27)
i∈I j∈J
x∈Xku
Pr

mb
and the predicted MB filter density fk|k−1 (Y ) ∼=
X j
X Y
f mbm Xkd = fCj X C ,



WA,k  

j,i
A∈Aj
, pj,i
U
X C =X d C∈A
C∈A rk|k−1 k|k−1 j
, and the backward smoothed
(24) i∈I j∈J

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
   
mb
density fk|l (Y ) ∼
=
j,i
rk|l , pj,i from l to k (k < l). forward PMBM filter recursion can refer to [41], [43]. Here
k|l

d
i∈Ij j∈J is omitted. We give a detailed GGIW implementation of
Then the backward smoothing recursion from time l to MB backward smoothing recursion based on the following
k − 1 is as follows  assumptions in this section.

we

1) The survival probability is state independent, i.e.,
 
mb ∼
fk−1|l = j,i j,i
rk−1|l , pk−1|l , (28)
j i∈I j∈J
pS,k|k−1 (x) = pS,k|k−1 .
2) The density of MB  with parameters at time
i.e., 
j,i

 Y rj,i pj,i (x) k − 1 is rk−1 , pj,i
k−1 , where pj,ik−1 (x) =
Y k−1|l k−1|l i∈Ij j∈J
mb j,i
fk−1|l (X) = 1− rk−1|l ·  , (29)  
GGIW x; aj,i j,i j,i j,i j,i j,i

j,i
1 − r k−1 , bk−1 , mk−1 , Pk−1 , vk−1 , Vk−1 .

ie
j∈J i∈Ij
k−1|l
3) The
 predicted density  of MB with parame-
where j,i


j,i

j,i
 ters is rk|k−1 , pj,i
k|k−1 , where pj,i k|k−1 (x) =
1 − rk−1 1 − rk|l i∈Ij j∈J

ev
j,i
rk−1|l =1− , (30)
 
j,i
1 − rk|k−1 GGIW x; aj,i , bj,i
k|k−1 k|k−1 , m j,i
k|k−1 , P j,i
, v j,i
k|k−1 k|k−1 , V j,i
k|k−1 .
4) The smoothed density  of MB  from l to k (k < l)
pj,i
k−1|l(x) = with parameters is j,i j,i
rk|l , pk|l , where pj,i
k|l (y) =
 i∈Ij j∈J
R fk|k−1 (y|x)pk|l (y) 

r
 
pk−1 (x) αS,k|l (x) + βS,k|l (x) dy GGIW y; aj,i j,i j,i j,i j,i j,i
k|l , bk|l , mk|l , Pk|l , vk|l , Vk|l .
pk|k−1 (y)
 R fk|k−1 (y|x)pk|l (y)  ,
Then the GGIW implementation of MB backward
R
pk−1 (x) αS,k|l (x) + βS,k|l (x) pk|k−1 (y) dy dx
er smoothing recursion step is given in Sec.IV B.
(31)
  4.1. Single extended target model
j,i

1 − rk|l 1 − pS,k|k−1 (x) The hybrid algorithm requires a single extended target
αS,k|l (x) ∼
= j,i
, (32)
1− rk|k−1 model. One of the most popular single extended target
pe
model is random matrix model due to its practicability
j,i
rk|l pS,k|k−1 (x) and universality [44, 45, 46], also known as the GGIW
βS,k|l (x) ∼
= . (33)
j,i
rk|k−1 model, in which the target shape is approximated as an
ellipse. In this section we give a brief review of random
Proof. See Appendix. matrix model, a longer review can be found in [7, Sec.III
A Bernoulli component corresponds to a target, multi- A].
Bernoulli components correspond to multiple targets. In the random matrix model, the extended target state
ot

Compare with the MB filter, the MBM filter performs well xk is the combination of a vector ξk ∈ Rnx representing
in terms of the tracking accuracy, but suffers a higher com- the target kinematic state, a matrix χk ∈ Sd (d is the
putational cost. If a single MB is used to represent targets dimension of the extent, typically d = 2 or d = 3) repre-
that have been detected, the MBM filter reduces to an senting the target extent state and a scalar γk being the
tn

MB filter. The proposed hybrid algorithm is able to track measurement model Poisson rate, i.e., xk = {ξk , χk , γk }.
multiple targets due to its utilization of PMBM RFS (see The motion models are given by
Sec.II C4) to model the states of multiple targets. The ξk = (ξk−1 ) + ωk−1 , (34a)
PMBM filter is an MTT conjugate prior, which is cen- χk = M (ξk−1 ) χk−1 M (ξk−1 ) ,
>
(34b)
tral in Bayesian theory, while the new method proposed
γk = γk−1 , (34c)
rin

in this paper is a PMBM approximation of generic multi-

target Bayesian forward–backward smoother. Compared
where F (·) is a motion model, ωk−1 is a zero mean Gaus-
with the PMBM filter, the proposed hybrid algorithm can
sian noise and M (·) is a transformation matrix. The mea-
achieve better tracking performance since more measure-
surement likelihood for a single measurement z is
ments are used to produce delayed estimations, which will
be demonstrated in a simulation study. φ (zk |xk ) = N (zk ; Hk ξk , σχk + Rk ) , (35)
ep

4. GGIW IMPLEMENTATION OF THE HY- where N (·) is the Gaussian distribution, along with Hk
BRID ALGORITHM is the measurement model, σ is a scaling factor and Rk is
In this work, the hybrid algorithm involved multiple the covariance of a zero mean Gaussian noise. The single
integrals is computationally intractable in general. Con- target density for the PPP model (10) with single measure-
Pr

sequently, the closed-form solution is given via the GGIW ment likelihood (35) is a product of Gamma, Gaussian and
implementation. The hybrid algorithm consists of the inverse-Wishart distributions [10],[45]
forward PMBM filter recursion and the MB backward f (X) = GAM (γ; a, b) N (ξ; m, P ) IWd (χ; v, V )
smoothing recursion. The GGIW implementation of the ∼
= GGIW (X; ζ) , (36)

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
 where GAM (γ; a, b) is the Gamma distribution defined on using a Bayes-optimal multi-target state estimator [11]. In

d
γ > 0 with shape parameter a and inverse scale parameter fact, we utilize the PMBM forward-backward smoothing
b, IWd (χ; v, V ) is the inverse-Wishart distribution defined recursion to compute the multi-target
 smootheddensity
on positive definite d × d matrix with degree of freedom v

j,i j,i
from l to k − 1 (k < l), i.e., rk−1|l , pk−1|l . In

we
and d × d scale matrix V and ζ = (a, b, m, P, v, V ) is the i∈Ij j∈J
GGIW density parameters. the case of the marginal multi-target (MaM) estimator, a
j,i
4.2. GGIW implementation of MB backward smoothing target can be declared to exist if rk−1|l > 1/2. Given this,
recursion MaM
the estimated state is X̂k−1|l = arg sup pj,i
k−1|l (X)[10].
X
Given the above four assumptions and a set of mea-
The pseudo-code of the hybrid algorithm is presented
surements Zk and standard multi-target system models

ie
in Algorithm.
(see Sec.II B), then the GGIW implementation of MB
backward smoothing  recursion from lto k − 1 (k < l) is
j,i
 Algorithm: Pseudo code for the hybrid algorithm
parameterized by rk−1|l , pj,i
k−1|l
j,i
, where rk−1|l

ev
j i∈I j∈J 1. Input: a posterior PMBM density
( )
refer to (30).    
From (31), we have 2. fk−1 ∼
= λu
k−1 , j
Wk−1 j,i
, rk−1 , pj,i
k−1 ,
i∈Ij j∈J
pj,i
k−1|l (x) =

R fk|k−1 (y|x)GGIW (y;ζk|l )
 3. and measurement set Z.
GGIW (x; ζk−1 ) αS,k|l + βS,k|l dy

r
GGIW (y;ζk|k−1 ) 4. PMBM forward recursion:
  ,
R fk|k−1 (y|x)GGIW (y;ζk|l ) j j,i
Initialize λu0|0 , W0|0 , pj,i
R
GGIW (x; ζk−1 ) αS,k|l + βS,k|l GGIW (y;ζk|k−1 )
dy dx 5. , r0|0 0|0
6. for k = 1, . . . , l (l > k) do
where 
j,i
1 − rk|t

1 − pS,k|k−1


er 7. j
compute λuk|k−1 , Wk|k−1 j,i
, rk|k−1 , pj,i
k|k−1

αS,k|l = , (37a) 8. according to (2)

j,i
1 − rk|k−1 j j,i
9. compute λuk|k , Wk|k , rk|k , pj,i
pe
k|k
j,i
rk|t pS,k|k−1 10. according to (3)
βS,k|l = j,i
, (37b)
rk|k−1 11. end for
 
ζk−1 = aj,i j,i j,i j,i j,i j,i 12. MB backward recursion:
k−1 , bk−1 , mk−1 , Pk−1 , vk−1 , Vk−1 (37c) ,
j,i
  13. Initialize rk|l , pj,i
k|l
ζk|k−1 = aj,i , bj,i
, mj,i
, P j,i
, v j,i
, V j,i
k|k−1 ,
k|k−1 k|k−1 k|k−1 k|k−1 k|k−1 14. for k = l − 1, . . . , 1 (l > k) do
ot

(37d) j,i
  15. compute rk−1|l , pj,i
k−1|l
ζk|l = aj,i j,i j,i j,i j,i j,i
k|l , bk|l , mk|l , Pk|l , vk|l , Vk|l . (37e) 16. according to (4)
17. end for
tn

4.3. Complexity reduction and state estimation

18. Normalize the weights of global hypothesis
4.3.1. Complexity reduction
In the recursive process, we have a GGIW density for 19. Perform State Estimation (see Sec.IV C2)
Poisson and Bernoulli components. In fact, they grow ex- MaM
20. Output: X̂k−1|l = arg sup pj,i
k−1|l (X) .
X
ponentially over time and lead to higher computational
rin

cost of the hybrid algorithm in subsequent time steps.

The general methods of complexity reduction, including 5. SIMULATION RESULTS
pruning , merging , can refer to [40], [41] for more de- In this section, we evaluate the performance of the hy-
tails. Furthermore, a more useful reduction method is brid algorithm in a scenario over a surveillance area with
done by recycling Bernoulli components which are less coalescence, where there are four targets doing random and
than a fixed threshold and approximates their distribu- irregular movement with a constant velocity. The ground
ep

tion as PPP, which are incorporated into the PPP intensity truth trajectories are shown in Fig. 3. The maximum
represented undetected targets in the PMBM RFS frame- number of global hypotheses we considered is Nh = 200.
work. For more details about recycling algorithm, please At each update step, we use ellipsoidal gating with thresh-
refer to [42], [47]. old ΓT = 10 to consider only the relevant measurements
4.3.2. State estimation for each single trajectory hypothesis. In addition, for each
Pr

In this paper, we focus on the ETT and information of global hypothesis, we use Murty’s algorithm to select the
interest—the number of extended targets, their positions, M -best new global hypotheses without having to evalu-
velocities, target extent, trajectories, and so on—can be ate the weights of all the new global hypotheses . For
extracted from the posterior distribution, e.g., fk|k (X|Zk ), more details about Murty’s algorithm, please refer to [32].

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
 5000 100 800

d
PMBM smoother
PMB smoother 600
y position (m)

MBM smoother
PMBM filter

Loc.
Tot.
0 50 400

we
200

-5000 0 0
-4000 -3000 -2000 -1000 0 1000 2000 3000 4000 0 20 40 60 80 100 0 20 40 60 80 100
x position (m) Time step Time step

5000 2000 3000

y position (m)

1500

ie
2000

Mis.

Fal.
0 1000
1000
500

ev
-5000 0 0
-4000 -3000 -2000 -1000 0 1000 2000 3000 4000 0 20 40 60 80 100 0 20 40 60 80 100
x position (m) Time step Time step

Figure 3: Top: The ground truth trajectories of four targets are Figure 4: Decomposition of GOSPA errors of PMBM filter, PMBM
shown in four different colors, their initial positions are marked with smoother and variant smoothers. They are total error (Tot.), local-

r
a circle, and final positions with a triangle. Bottom: The generated ization error (Loc.), missed error (Mis.) and false error (Fal.).
measurement data (black crosses) are shown in per MC.

matrix, they aregiven by respectively

Once the new global hypotheses are formed, pruning is
performed to keep the Nh global hypotheses with highest
er H=
1 0 0 0
0 1 0 0

, R = σε

2 1
0
0
1

,
weights. Therefore, Bernoulli components with existence
probability smaller than ΓB = 10−3 , PPP components where σε = 1m denotes the standard deviation of mea-
pe
with existence probability smaller than ΓP = 10−3 and surement noise.
GGIW components with weights smaller than ΓG = 10−3 The Optimal Sub-pattern Assignment (OSPA) [48] and
are both pruned in the PMBM filter recursion or MB generalized GOSPA metric [49] are used for performance
smoothing recursion. In addition, the recycling method assessments. The cut-off parameter and the order param-
is applied to Bernoulli components with existence proba- eter are set to c = 20, p = 2, respectively.
bility smaller than ΓR = 10−2 . In the simulation exper- The decomposition of GOSPA errors, OSPA error and
cardinality estimate of PMBM Filter, Smoother and vari-
ot

iment presented here, we focus on comparing the hybrid

algorithm (Hereinafter referred to as “PMBM Smoother”.) ant Smoothers are shown in Figs. 4, 5, meanwhile, along
with the standard PMBM filter (Hereinafter referred to as with the mean computational time per MC are presented
“PMBM Filter”.) and other variant smoothers. in Table 2.
It can be seen that the PMBM Smoother has the bet-
tn

For each filter, 100 Monte Carlo (MC) runs were per-
formed. The four targets move following a constant ve- ter estimation performance because of the minimum total
locity model. The process and measurement noises are all error and stable cardinality estimate and more measure-
zero-mean Gaussian with standard deviation 0.1 for each ments are used to produce more effective estimations by
dimension. The Poisson clutter intensity is λC = 10 and delaying the decision time (k) and using data at a later
the target detection probability is PD = 0.90, target sur- time l (l > k). . However, “Smoother” is generally a com-
rin

vival probability is PS = 0.90. The Poisson birth intensity putationally expensive operation since the mean compu-
is a GGIW mixture tational time of “Smoother” is slower than “Filter”, the
X4 reason is that “Smoother” needs to perform the third step
λbk (x) = rkb GGIW aik , bik , mik , Pki , vki , Vki ,


“smoothing” after “prediction” and “filtering”, while “Fil-
i=1 ter” does not. In other words, the “Smoother” density
ep

needs to be computed in reverse order from k − 1 to 0 with

where rkb = 0.01 is birth probability, and a = 30, b = 5, the following MB recursive density after “update” using
 2
> > measurements at a current time k.
m = [1/3; 1/3; 1/3] , P = diag [1, 1, 1, 1] , v = 12,

>
 In addition, since the PMBM conjugate prior has a
V = diag [12, 12] . more efficient hypothesis structure than the MBM con-
Let ξk = [pk , vk ] ∈ R4 represents the kinematic state, jugate prior, the filters (PMBM and PMB) with Poisson
Pr

where pk ∈ R2 is the position of extended target, vk ∈ birth model outperform their counterparts (MBM) with
R2 is the velocity. l (z|x) = N (z; Hx, R), where H is MB birth model in terms of GOSPA errors. The PMB
observation matrix and R is measurement noise covariance filter with variational MB approximation not only has a

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
 The newborntargets can be denoted by an MB RFS
5

d
True
 
PMBM smoother
mb
fB,k|k−1 ∼
=
j,i
rB,k|k−1 , pj,i . The multi-
4 B,k|k−1
PMB smoother i∈Ij
Cardinality

MBM smoother j∈J

3 PMBM filter target transition density fk|k−1 (Y |X) can be denoted by

we
4

2
(6). The backward smoothing recursion (4) can be refor-
44.5 45 45.5
mulated as
1 mb
0 10 20 30 40 50 60 70 80 90 100 fk−1|l (X) =
Time step Z
mb
(X) fB,k|k−1 Y +


200 fk−1
150
OSPA Errors

mb

ie

fk|l (Y )
100 fS,k|k−1 Y − |X mb
δY
fk|k−1 (Y )
50
−
Z
−
mb

fk|l (Y − )
mb
fS,k|k−1 Y |X δY −

ev
0
0 20 40 60 80 100
= fk−1 (X) mb −
Time step
fk|k−1 (Y − )
| {z }
Survived targets
+
mb
Z
+

fk|l (Y + )
Figure 5: Cardinality estimate, mean OSPA error of PMBM filter,
· fB,k|k−1 Y mb + δY + , (39)
PMBM smoother and variant smoothers. fk|k−1 (Y + )

r
| {z }
very close estimation performance to the PMBM filter in Born targets

the scenario, but also significantly faster than the PMBM er where
filter. The reason is that to approximate the PMBM pos- −

mb+
mb mb
Y − fk|l Y+ ,


terior density as a PMB after the update step via the fil- fk|l (Y ) = fk|l (40a)
mb−

mb−
tering recursion is performed, which can be with far fewer mb
Y − fk|l Y+ .


fk|k−1 (Y ) = fk|k−1 (40b)
Bernoulli components and lower computational complex-
pe
ity. As fB,k|k−1 (Y + ) = fk|k−1
mb +
(Y + ), the second part of (39)
6. CONCLUSION is equal to 1, i.e.,
mb+
In this paper, we have derived a hybrid algorithm un-
Z

fk|l (Y + )
der the RFS framework and has the advantages for main- fB,k|k−1 Y + mb+ δY + =
fk|k−1 (Y + )
taining the tracking accuracy of different locations. We Z
also give an implementation by resorting to the GGIW mb+
Y + δY + = 1.


fk|l (41)
ot

technology for the proposed method. Both numerical and

simulation results have demonstrated the superior perfor-
From (7, 32, 33, 38, 41), (39) is rewritten as follows
mance of the hybrid algorithm in terms of GOSPA (OSPA) mb
errors and tracking performance as compared to the stan- fk−1|l (X) =
tn

mb−
dard PMBM filter and variant smoothers. Z
−

fk|l (Y − )
mb
fk−1 (X) fS,k|k−1 Y |X mb− δY −
fk|k−1 (Y − )
APPENDIX
j,i !|J|−1 
1 − rk|l X
PROOF OF PROPOSITION mb
= fk−1 (X) j,i
1 − pj,i
S,k|k−1
1 − rk|k−1
rin

The set integral of the backward smoothing recursion   

(4) can be divided into two parts (see (39)): One is the
Z  pj,i j,i
S,k|k−1 (x) 1 − rk|k−1 rk|l
j,i

survived targets, the other is the newborn targets. The 

 1 − pj,i

j,i

j,i
(x) 1 − rk|l rk|k−1
second part (i.e., the newborn targets) is equal to 1 as S,k|k−1
in (41). That is mean to the newborn targets can not Y −
fk|k−1 (y|x) pk|l (y)
ep

transmit to the time before birth. As a consequence, the · δY −

pk|k−1 (y)
set integral of (4) equals to the set integral of the first part
j,i !|J|−1−|X|
(i.e., the survived targets). Lastly, we prove that the set 1 − rk|l
X
mb
integral of the first part is an MB. = fk−1 (X) j,i
αS,k|l (x)
We assume that fk−1 mb mb
(X), fk|k−1 (Y ) and fk|lmb
(Y ) are 1− rk|k−1
Pr

mb
all MB distributions. fk|l (Y ) is described as follows Z  |Y − |  Y −
βS,k|l (x) fk|k−1 (y|x) pk|l (y)
· δY − .
X Y  Y rj,i pj,i (y) αS,k|l (x) pk|k−1 (y)
mb j,i k|l k|l
fk|l (Y ) = 1 − rk|l j,i
. (38) (42)
]Y j =Y i j∈J i∈Ij
1 − rk|l

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
 Table 2: Mean GOSPA Errors, OSPA Error and Computational Time (s) of PMBM Filter, Smoother and variant Smoothers

d
GOSPA-Tot. Loc. Mis. Fal. OSPA Time (s)

we
PMBM Smoother 16.2193 539.4699 294.3826 454.6922 51.5038 8.9817
PMB Smoother 21.1460 561.8160 321.8539 528.0292 55.9602 8.4199
MBM Smoother 24.2360 582.1893 353.6995 624.4157 67.0416 9.5622
PMBM Filter 28.2051 604.1475 400.3673 771.1817 78.8477 7.3015

ie
j,i !|J|−1−|X|
Applying Lemma 3 of [16] and the power-functional 1 − rk|l
mb mb
identity in Section 3.7 of [11], we have fk−1|l (X) = fk−1 (X) j,i
1 − rk|k−1
j,i !|J|−1−|X| Y 
βS,k|l (x) fk|k−1 (y|x) pk|l (y)
Z
1 − rk|l

ev

X
mb
fk−1|l (X) = mb
fk−1 (X) αS,k|l (x) αS,k|l (x) + dy
j,i
1 − rk|k−1 pk|k−1 (y)
y∈Y
  
Z  |Y − |  Y − j,i
1 − rk−1 j,i
1 − rk|l
βS,k|l (x) fk|k−1 (y|x) pk|l (y)
δY −
Y
=   ·
αS,k|l (x) pk|k−1 (y) j,i
1 − rk|k−1
j∈|J|−1−|X|

r
j,i !|J|−1−|X|
1 − rk|l
X Y n j,i j,i
mb
= fk−1 (X) αS,k|l (x) rk−1 pk−1 (x) · αS,k|l (x)
j,i
1− rk|k−1 er i∈|X|

βS,k|l (x) fk|k−1 (y|x) pk|l (y)
Z
βS,k|l (x) fk|k−1 (y|x) pk|l (y)
X Y Z
dy + dy . (44)
−
αS,k|l (x) pk|k−1 (y) pk|k−1 (y)
Y ⊂Y y∈Y
j,i !|J|−1−|X| Let
1 − rk|l
X
pe
mb j,i
= fk−1 (X) j,i
αS,k|l (x) mb
fk−1|l (x) = rk−1|l · pj,i
k−1|l (x)
1 − rk|k−1
j,i
Y  = rk−1 · pj,i
k−1 (x) αS,k|l (x) +
βS,k|l (x) fk|k−1 (y|x) pk|l (y)
Z
1+ dy Z β j,i !
αS,k|l (x) pk|k−1 (y) S,k|l (x) fk|k−1 (y|x) pk|l (y)
y∈Y dy . (45)
j,i !|J|−1−|X|
pj,i
k|k−1 (y)
mb
1 − rk|l
= fk−1 (X)
ot

j,i We have
1 − rk|k−1 Z
j,i mb
Y Z
βS,k|l (x) fk|k−1 (y|x) pk|l (y)
 rk−1|l = fk−1|l (x) dx. (46)
αS,k|l (x) + dy .
pk|k−1 (y)
y∈Y mb mb
fk−1|l (x) fk−1|l (x)
tn

(43) pj,i
k−1|l (x) = j,i
=R mb (x) dx
. (47)
rk−1|l fk−1|l
Substituting
Y   Y rj,i pj,i (x) Substituting (30),(45),(46),(47) into (44) yields
mb j,i k−1 k−1
fk−1 (X) = 1 − rk−1 j,i
mb
fk−1|l (X) =
j∈|J|−1 i∈|X|
1 − rk−1
rin

Y   Y  
j,i j,i
1 − rk−1|l · rk−1|l pj,i
k−1|l (x)
into (43) yields j∈|J|−1−|X| i∈|X|
j,i
Y  j,i
 Y rk−1|l pj,i
k−1|l (x)
= 1 − rk−1|l ·   . (48)
j,i
j∈|J|−1 i∈|X| 1 − rk−1|l
ep

From (48), we get the parameter set (28) which indi-

j,i
mb
cates that fk−1|l (X) is an MB where rk−1|l and pj,i
k−1|l (x)
are given in (30) and (31), respectively. So far, (29) has
been proved.
Pr

We can obtain (31) by substituting (45) into (47).

From (46), we can get (30) and the proof is given as fol-
lows.
Firstly, we have

10

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
 Z
j,i j,i
rk|k−1 = rk−1 pj,i j,i
S,k|k−1 (x) pk−1 (x) dx. (49) ternational Conference on Advanced Video and Signal Based
Surveillance (AVSS), IEEE, 2016, pp. 23–29.

d
[4] L. Gao, G. Battistelli, L. Chisci, Random-finite-set-based dis-
tributed multirobot slam, IEEE Transactions on Robotics 36 (6)
j,i
pj,i (x) pj,i
R
rk−1 k−1 S,k|k−1 (x) fk|k−1 (y|x)dx (2020) 1758–1777.

we
pj,i
k|k−1 (y) = j,i
. [5] S. Scheidegger, J. Benjaminsson, E. Rosenberg, A. Krishnan,
rk|k−1 K. Granström, Mono-camera 3d multi-object tracking using
(50) deep learning detections and pmbm filtering, in: 2018 IEEE
Intelligent Vehicles Symposium (IV), IEEE, 2018, pp. 433–440.
[6] K. Granström, A. Natale, P. Braca, G. Ludeno, F. Serafino,
From (45),(46),
Z we get Gamma gaussian inverse wishart probability hypothesis density
j,i mb for extended target tracking using x-band marine radar data,
rk−1|l = fk−1|l (x) dx IEEE Transactions on Geoscience and Remote Sensing 53 (12)

ie
Z (2015) 6617–6631.
j,i
= αS,k|l (x) rk−1 pj,i
k−1 (x) dx+
[7] K. Granström, M. Baum, S. Reuter, Extended object tracking:
Introduction, overview and applications, Journal of Advances
Z Z rj,i pj,i (x) β j,i in Information Fusion 12 (2016) 139–174.
S,k|l (x) fk|k−1 (y|x) pk|l (y)

ev
k−1 k−1 [8] T. Fortmann, Y. Bar-Shalom, M. Scheffe, Sonar tracking of mul-
dydx, tiple targets using joint probabilistic data association, IEEE
pj,i
k|k−1 (y) journal of Oceanic Engineering 8 (3) (1983) 173–184.
(51) [9] C.-Y. Chong, S. Mori, D. B. Reid, Forty years of multiple hy-
pothesis tracking-a review of key developments, in: 2018 21st
where the first term on the right of (51) is written as fol- International Conference on Information Fusion (FUSION),
IEEE, 2018, pp. 452–459.

r
lowsZ [10] R. P. Mahler, Statistical multisource-multitarget information
j,i
αS,k|l (x) rk−1 pj,i
k−1 (x) dx
fusion, Vol. 685, Artech House Norwood, MA, USA, 2007.
er [11] R. P. Mahler, Advances in statistical multisource-multitarget

j,i

j,i Z information fusion, Artech House, 2014.
1 − rk|l rk−1   [12] R. P. Mahler, Multitarget bayes filtering via first-order multitar-
j,i
= j,i
1 − p S,k|k−1 (x) pj,i
k−1 (x) dx get moments, IEEE Transactions on Aerospace and Electronic
1 − rk|k−1 systems 39 (4) (2003) 1152–1178.
j,i  [13] R. P. Mahler, Phd filters of higher order in target number, IEEE
1 − rk|l  Transactions on Aerospace and Electronic systems 43 (4) (2007)
pe
j,i j,i
= j,i
rk−1 − rk|k−1 . (52) 1523–1543.
1 − rk|k−1 [14] B.-T. Vo, B.-N. Vo, A. Cantoni, The cardinality balanced multi-
target multi-bernoulli filter and its implementations, IEEE
The second term on the right of (51) is written as follows Transactions on Signal Processing 57 (2) (2008) 409–423.
Z Z rj,i pj,i (x) β j,i [15] S. Reuter, B.-T. Vo, B.-N. Vo, K. Dietmayer, The labeled multi-
k−1 k−1 S,k|l (x) fk|k−1 (y|x) pk|l (y) bernoulli filter, IEEE Transactions on Signal Processing 62 (12)
dydx (2014) 3246–3260.
pj,i
k|k−1 (y) [16] B.-T. Vo, B.-N. Vo, Labeled random finite sets and multi-
ot

Z R rj,i pj,i (x) pj,i j,i j,i object conjugate priors, IEEE Transactions on Signal Process-
k−1 k−1 S,k|k−1 fk|k−1 (y|x)dx rk|l pk|l (y) ing 61 (13) (2013) 3460–3475.
= j,i
· j,i dy
rk|k−1 pk|k−1 (y) [17] J. L. Williams, Marginal multi-bernoulli filters: Rfs derivation
Z of mht, jipda, and association-based member, IEEE Transac-
j,i j,i j,i tions on Aerospace and Electronic Systems 51 (3) (2015) 1664–
= rk|l pk|l (y) dy = rk|l . (53)
tn

1687.
[18] Á. F. Garcı́a-Fernández, J. L. Williams, K. Granström,
L. Svensson, Poisson multi-bernoulli mixture filter: Di-
Lastly, combining (51-53) yields rect derivation and implementation, IEEE Transactions on
j,i 
j,i
1 − rk|l j,i j,i

j,i
Aerospace and Electronic Systems 54 (4) (2018) 1883–1901.
rk−1|l = j,i
rk−1 − rk|k−1 + rk|l [19] A. F. Karr, Markov chains: Theory and applications (dean l.
1 − rk|k−1 isaacson and richard w. madsen), SIAM Review 20 (3) (1978)
rin

   606.
j,i j,i
1 − rk−1 1 − rk|l [20] B.-N. Vo, B.-T. Vo, R. P. Mahler, Closed-form solutions to
=1− j,i
.  (54) forward–backward smoothing, IEEE Transactions on Signal
1 − rk|k−1 Processing 60 (1) (2011) 2–17.
[21] A. C. Harvey, Forecasting, structural time series models and the
kalman filter (1990).
References [22] J. Meditch, A survey of data smoothing for linear and nonlinear
ep

dynamic systems, Automatica 9 (2) (1973) 151–162.

[1] B.-N. Vo, M. Mallick, Y. Bar-Shalom, S. Coraluppi, R. Osborne, [23] B.-T. Vo, D. Clark, B.-N. Vo, B. Ristic, Bernoulli forward-
R. Mahler, B.-t. Vo, Multitarget tracking, Wiley encyclopedia backward smoothing for joint target detection and tracking,
of electrical and electronics engineering (2015) (2015). IEEE Transactions on Signal Processing 59 (9) (2011) 4473–
[2] S. Pang, D. Morris, H. Radha, 3d multi-object tracking using 4477.
random finite set-based multiple measurement models filtering [24] S. Wong, B. T. Vo, F. Papi, Bernoulli forward-backward
Pr

(rfs-m 3) for autonomous vehicles, in: 2021 IEEE International smoothing for track-before-detect, IEEE Signal Processing Let-
Conference on Robotics and Automation (ICRA), IEEE, 2021, ters 21 (6) (2014) 727–731.
pp. 13701–13707. [25] N. Nadarajah, T. Kirubarajan, T. Lang, M. McDonald,
[3] T. L. A. Nguyen, F. Bremond, J. Trojanova, Multi-object track- K. Punithakumar, Multitarget tracking using probability hy-
ing of pedestrian driven by context, in: 2016 13th IEEE In- pothesis density smoothing, IEEE Transactions on Aerospace

11

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720
 and Electronic Systems 47 (4) (2011) 2344–2360. [45] M. Feldmann, D. Fränken, W. Koch, Tracking of extended ob-
[26] R. P. Mahler, B.-T. Vo, B.-N. Vo, Forward-backward probability jects and group targets using random matrices, IEEE Transac-

d
hypothesis density smoothing, IEEE Transactions on Aerospace tions on Signal Processing 59 (4) (2010) 1409–1420.
and Electronic Systems 48 (1) (2012) 707–728. [46] K. Granström, U. Orguner, A phd filter for tracking multiple
[27] D. E. Clark, First-moment multi-object forward-backward extended targets using random matrices, IEEE Transactions on

we
smoothing, in: 2010 13th International Conference on Infor- Signal Processing 60 (11) (2012) 5657–5671.
mation Fusion (FUSION), IEEE, 2010, pp. 1–6. [47] J. L. Williams, Hybrid poisson and multi-bernoulli filters, in:
[28] B.-N. Vo, B.-T. Vo, R. P. Mahler, A closed form solution to the 2012 15th International Conference on Information Fusion (FU-
probability hypothesis density smoother, in: 2010 13th Inter- SION), IEEE, 2012, pp. 1103–1110.
national Conference on Information Fusion (FUSION), IEEE, [48] D. Schuhmacher, B.-T. Vo, B.-N. Vo, A consistent metric for
2010, pp. 1–8. performance evaluation of multi-object filters, IEEE transac-
[29] S. Nagappa, E. D. Delande, D. E. Clark, J. Houssineau, A tions on signal processing 56 (8) (2008) 3447–3457.
tractable forward–backward cphd smoother, IEEE Transactions [49] A. S. Rahmathullah, Á. F. Garcı́a-Fernández, L. Svensson, Gen-

ie
on Aerospace and Electronic Systems 53 (1) (2017) 201–217. eralized optimal sub-pattern assignment metric, in: 2017 20th
[30] Y. Xia, L. Svensson, Á. F. Garcı́a-Fernández, K. Granström, International Conference on Information Fusion (FUSION),
J. L. Williams, Backward simulation for sets of trajectories, IEEE, 2017, pp. 1–8.
in: 2020 23rd International Conference on Information Fusion

ev
(FUSION), IEEE, 2020, pp. 1–8.
[31] M. Beard, B. T. Vo, B.-N. Vo, Generalised labelled multi-
bernoulli forward-backward smoothing, in: 2016 19th Inter-
national Conference on Information Fusion (FUSION), IEEE,
2016, pp. 688–694.
[32] Á. F. Garcı́a-Fernández, L. Svensson, M. R. Morelande, Multi-
ple target tracking based on sets of trajectories, IEEE Transac-

r
tions on Aerospace and Electronic Systems 56 (3) (2019) 1685–
1707.
[33] R. Liu, H. Fan, T. Li, H. Xiao, A computationally efficient
er
labeled multi-bernoulli smoother for multi-target tracking, Sen-
sors 19 (19) (2019) 4226.
[34] B.-N. Vo, B.-T. Vo, A multi-scan labeled random finite set
model for multi-object state estimation, IEEE Transactions on
signal processing 67 (19) (2019) 4948–4963.
[35] G. Liang, Q. Li, B. Qi, L. Qiu, Multitarget tracking using
pe
one time step lagged delta-generalized labeled multi-bernoulli
smoothing, IEEE Access 8 (2020) 28242–28256.
[36] B. Yu, E. Ye, Track-before-detect labeled multi-bernoulli
smoothing for multiple extended objects, in: 2020 23rd Inter-
national Conference on Information Fusion (FUSION), IEEE,
2020, pp. 1–8.
[37] T. Li, H. Chen, S. Sun, J. M. Corchado, Joint smoothing and
tracking based on continuous-time target trajectory function
ot

fitting, IEEE Transactions on Automation Science and Engi-

neering 16 (3) (2018) 1476–1483.
[38] T. Li, X. Wang, Y. Liang, J. Yan, H. Fan, A track-oriented
approach to target tracking with random finite set observations,
in: 2019 International Conference on Control, Automation and
tn

Information Sciences (ICCAIS), IEEE, 2019, pp. 1–6.

[39] Y. Xia, K. Granström, L. Svensson, Á. F. Garcı́a-Fernández,
J. L. Williams, Extended target poisson multi-bernoulli mixture
trackers based on sets of trajectories, in: 2019 22th International
Conference on Information Fusion (FUSION), IEEE, 2019, pp.
1–8.
[40] Y. Xia, K. Granström, L. Svensson, M. Fatemi, Á. F. Garcı́a-
rin

Fernández, J. L. Williams, Poisson multi-bernoulli approxima-

tions for multiple extended object filtering, IEEE Transactions
on Aerospace and Electronic Systems 58 (2) (2022) 890–906.
[41] K. Granström, M. Fatemi, L. Svensson, Poisson multi-bernoulli
mixture conjugate prior for multiple extended target filtering,
IEEE Transactions on Aerospace and Electronic Systems 56 (1)
ep

(2019) 208–225.
[42] X. Xie, Y. Wang, Analysis of recycling performance in pois-
son multi-bernoulli mixture filters, in: 2021 24th International
Conference on Information Fusion (FUSION), IEEE, 2021, pp.
1–7.
[43] K. Granström, M. Fatemi, L. Svensson, Gamma gaussian
inverse-wishart poisson multi-bernoulli filter for extended target
Pr

tracking, in: 2016 19th International Conference on Information

Fusion (FUSION), IEEE, 2016, pp. 893–900.
[44] J. W. Koch, Bayesian approach to extended object and clus-
ter tracking using random matrices, IEEE Transactions on
Aerospace and Electronic Systems 44 (3) (2008) 1042–1059.

12

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4327720