Signal Processing 151 (2018) 32–44

Contents lists available at ScienceDirect

Signal Processing
journal homepage: www.elsevier.com/locate/sigpro

Bearings-only multi-target tracking using an improved labeled

multi-Bernoulli filter
Yifan Xie, Taek Lyul Song∗
Department of Electronic Systems Engineering, Hanyang University, Ansan, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Article history: Most classical bearing-only target tracking algorithms model the measurement likelihood by one Gaus-
Received 9 December 2017 sian distribution. The effectiveness of one Gaussian distribution model relies heavily on the accuracy of
Revised 28 April 2018
the predicted target position. However, due to the high nonlinearity of the bearing-only measurement,
Accepted 30 April 2018
the predicted target position is mostly inaccurate before the target state observability is established. As a
Available online 1 May 2018
consequence, some classical nonlinear filters become not applicable for tracking bearing-only targets, es-
Keywords: pecially when the measurements of multiple targets and clutter are present. The published bearings-only
Bearing-only multiple-target tracking algorithms suffer from either the estimation inaccuracy or lack of track trajecto-
Multiple targets ries. Motivated by the problems mentioned above, we propose an improved labeled multi-Bernoulli filter
Gaussian mixture measurements for the goal of reducing estimation error under the premise that track trajectories are guaranteed. The
Labeled multi-Bernoulli proposed method divides the bearing measurement uncertainty into several measurement components
that the measurement likelihood can be approximated by a Gaussian mixture. By assigning each track a
unique label, the previous scan estimations and current scan measurements are associated and the track
trajectories become available. Simulation results show that the proposed method considerably reduces
estimation error. Further, various scenario parameters are investigated to validate the effectiveness of the
proposed method.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction points, while the particle filter (PF) [8] uses a large number of
weighted random (Monte Carlo) samples. When only the bearing
The problem of bearing-only tracking (BOT) [1] can be referred information is provided, the distance between a passive observer
to as target motion analysis (TMA) [2] and has been widely studied and a target is not available, which leads to the system state
for decades. This problem is crucial for a variety of surveillance observability problem for the BOT problem. Due to the unknown
systems such as mobile platforms equipped with passive sonar distance, the target of interest cannot be distinguished from the
that detects the radiated signals from moving objects or air- other targets of that direction (bearing). Thus, the observer must
crafts using an electronic warfare device [3,4]. The bearing-only ‘outmaneuver’ the target to satisfy the observability condition to
tracking problem is essential due to two features: measurement recognize the target of interest, i.e., the observer motion model
nonlinearity and lack of system observability [5]. The extended must be at least one derivative higher than that of the target [9].
Kalman filter (EKF) [6] has been generally applied for target For practical applications, a filter structure such as the EKF, the
tracking with nonlinear measurements. The EKF linearizes the UKF and the PF cannot be directly applied for tracking bearing-
measurement nonlinearity around the predicted target position only targets in cluttered environments. One has to resort to the
through first-order approximation, which holds when the pre- data association algorithms to account for the measurement origin
dicted target position is accurate enough but can introduce large uncertainty due to the presence of multiple targets and clutter
estimation errors or even divergence in the presence of inaccuracy measurements. In addition, due to the limited prior information
[7]. Instead of linearizing the measurement nonlinearity, the of the surveillance region, real target tracking systems usually
unscented Kalman filter (UKF) [7] utilizes the unscented transform utilize the received measurements for track initiation. The received
to sample and propagate the probability density function by sigma measurements can either originate from the true targets or clutter
such that both true tracks (initialized by target measurements)
and false tracks (initialized by clutter measurements) are gener-
∗
Corresponding author. ated. The false tracks that follow wrong or not existing targets are
E-mail addresses: [email protected] (Y. Xie), [email protected] (T.L.
subjected to termination. Then, a metric of evaluating track quality
Song).

https://doi.org/10.1016/j.sigpro.2018.04.027
0165-1684/© 2018 Elsevier B.V. All rights reserved.
 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44 33

Nomenclature yak Mean of Gaussian measurement compo-

nent a’s probability density function at
A. List of acronyms scan k
BOT Bearing-only tracking Rak Covariance of Gaussian measurement
TMA Target motion analysis component a’s probability density func-
EKF Extended Kalman filter tion at scan k
UKF Unscented Kalman filter γka Weight of Gaussian measurement com-
PF Particle filter ponent a at scan k
GMM Gaussian mixture measurement zk, i The ith measurement in Zk
GMM-ITS Gaussian mixture measurement integrated track p(x,k
s)
Sensor position in x axis
splitting Tk, i Rotation matrix of measurement zk, i
MTT Multi-target tracking λk Coordinate transformation factor at scan
MHT Multiple hypothesis tracking k
JIPDA Joint integrated probabilistic data association c Track component index
GMM-JITS Gaussian mixture measurement joint integrated Ck The number of track components at scan
track splitting k
HPRF High pulse repetition frequency ξ k (c) Track component probability
RFS Random finite set p(x, |c, Z) Probability density function of track
CBMeMBer Cardinality balanced multi-target multi-Bernoulli component c in track 
δ -GLMB δ -generalized labeled multi-Bernoulli xc Mean of track component c’s probability
LMB Labeled multi-Bernoulli density function
SLAM Simultaneous localization and mapping Pc Covariance of track component c’s
LMB-GMM Labeled multi-Bernoulli with Gaussian mixture probability density function
measurements X+ Predicted LMB RFS
SMC Sequential Monte Carlo W The survival LMB RFS
SIR Sequential importance resampling ω+,S (L ) Weight of the hypothesis that L is the
RPF Regularized particle filter label set of survival tracks
OSPA Optimal sub-pattern assignment ( )
r+ ,S
Predicted target existence probability of
LMB-EKF Labeled multi-Bernoulli with extended Kalman survival track with label 
filter p(+,S) Predicted probability density function of
LMB-UKF Labeled multi-Bernoulli with unscented Kalman survival track with label 
filter pS ( x, ) State-dependent target survival proba-
LMB-SMC Labeled multi-Bernoulli with sequential Monte bility
Carlo implementation x̄ Predicted single-target state
Y The newborn LMB RFS
B. List of symbols
ω B (L) Weight of the hypothesis that L is the
xk Single-target state at scan k
label set of newborn tracks
Xk Multi-target state at scan k
rB( ) Initial target existence probability of
Zk Multi-target observation at scan k
newborn track with label 
Zk Sequence of measurement sets up to
p(B ) Initial probability density function of
scan k
newborn track with label 
X Target state space (i )
X+ The ith group of the predicted LMB RFS
Z Measurement space
θ Measurements-to-tracks association
L Label space
θ ( ) The measurement index associated to
π k (Xk |Zk ) Multi-target state density Xk given
track  under θ
measurement sequence Zk
F (L ) The collection of finite subsets of L
r Target existence probability
The space of measurements-to-tracks
 Track label
association
p( x, ) Probability density function of track
ω (I+ ,θ ) Updated weight of the hypothesis that
with label 
the predicted track labels in set I+ are
x Labeled single-target state
associated with measurements under θ
X Labeled multi-target state
g(zk, i |x, ) The measurement likelihood of mea-
( X ) Distinct label indication to ensure labels
surement zk, i with respect to track 
in X are distinct
pD ( x, ) State-dependent target detection proba-
L (X ) The set of labels of X
bility
ω (L) Weight of the hypothesis that L is the
pG Gating probability that the true mea-
set of track labels
surement falls in the gate
σφ Sensor noise standard deviation
κ (zk, i ) Poisson clutter intensity at measurement
Ak The number of Gaussian measurement
zk, i
components generated by one bearing
p(θ ) (x, |Z k+1 ) Updated single-target probability density
measurement
function of track  under θ
a Gaussian measurement component index
p(θ ) (x, |c, a, Z k+1 ) Probability density function of the new
L̄k,a The mean range of interval a
track component generated by measure-
Lk, a The length of range interval a
34 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44

which inherits the advantages of CBMeMBer and δ -GLMB such as

ment component a and track component the tractable computational load and track labels, is considered
c under θ in this paper for tracking multiple targets using bearings-only
ξ k (c, a) Track component probability of (c, a) measurements.
The LMB filter and the GLMB filter have been applied to solve
the problems such as simultaneous localization and mapping
to discriminate the false tracks from the true tracks becomes (SLAM) [23], constrained sensor control [24] and tracking with
necessary. Musicki [10] proposes a Gaussian Mixture Measure- merged measurements [25]. Although [25] demonstrates its ef-
ment (GMM) method to approximate measurement likelihood fectiveness in handling the merged bearing-only measurements,
by a Gaussian mixture and shows better estimation accuracy it emphasizes the merged measurement problem whereas the
compared to the EKF and the UKF for bearing-only target tracking. estimation inaccuracy caused by the imprecise measurement like-
The Gaussian Mixture Measurements-Integrated Track Splitting lihood is overlooked. In this paper, an improved LMB filter called
(GMM-ITS) filter extends the applicability of GMM to track targets the Labeled Multi-Bernoulli with Gaussian Mixture Measurements
in cluttered environments by incorporating the target existence (LMB-GMM) filter is proposed to complement the existing litera-
probability as a track quality measure [10]. But, the GMM-ITS tures for the BOT problem. The incorporation of the GMM method
filter is only applicable for tracking a single target in clutter, the improves the estimation accuracy of the standard LMB filter for
possibility of multiple targets in one surveillance region is not bearings-only target tracking. The LMB framework in turn extends
considered. the applicability of the GMM method for the MTT problem under
Due to the presence of multiple targets, real tracking systems cluttered environments. The LMB-GMM filter is not a simple appli-
also require track labels and measurement-to-track association cation of two existing methods. Due to the fact that the standard
method. The classical methods for multi-target tracking (MTT) with LMB algorithm cannot operate recursively in a closed form, the
track labels include the Multiple Hypothesis Tracking (MHT)[11] LMB-GMM is implemented as a three-stage procedure when the
and the Joint Integrated Probabilistic Data Association (JIPDA) filter filter status is updated. The first stage transforms the predicted
[12]. The MHT and the JIPDA filter utilize an optimal approach to LMB to a representation of predicted δ -GLMB with compromising
enumerate all possible measurement-to-track associations under a single hypothesis component; the second stage updates the
the assumption that one target generates zero (miss detection) δ -GLMB with Gaussian Mixture Measurements (δ -GLMB-GMM);
or one measurement at each scan. Inspired by the work of joint the last stage transforms the updated δ -GLMB back into an up-
multi-target data association in the MHT and the JIPDA filter, the dated LMB form. Specially, one specific measurement-to-track
GMM-Joint Integrated Track Splitting (GMM-JITS) filter applies association can occur repeatedly such that the internal relation-
the GMM method to tracking multiple targets using a High Pulse ship between updated tracks is addressed by appropriate track
Repetition Frequency (HPRF) radar [13]. Due to the range ambi- component managements, which makes the LMB-GMM distinct
guity, the GMM-JITS filter in [13] approximates the likelihood of from the other GMM based approaches. In this paper, the classical
each ambiguous HPRF radar range measurement by a number nonlinear filters (EKF, UKF and PF) are also associated with the
of unambiguous Gaussian mixture measurement components. LMB algorithm for tracking performance studies.
However, the GMM method for tracking multiple bearings-only The rest of the paper is organized as follows. Background infor-
targets with track labels has not been investigated. Although mation of the multi-target Bayes filter, the RFS, and the labeled RFS
the GMM-JITS seems to be an alternative for the bearings-only are detailed in Section 2. A brief description of the GMM method
multi-target tracking problem, it usually suffers from computa- is shown in Section 3. The proposed LMB-GMM filter is intro-
tional intractability. The computational intractability indicates that duced in Section 4. Details of the LMB-GMM implementation are
problems can be solved in theory but requires significant compu- described in Section 5. Analyses for the computational complexity
tation time to produce useful solutions [14]. For scenarios where are discussed in Section 6. Simulations of the proposed method are
bearing measurements are distributed closely in measurement presented in Section 7, followed by the conclusions in Section 8.
space, the number of measurement-to-track associations increases
exponentially so that computational intractability occurs.
By modeling the set-valued target states based on the set- 2. Background
valued measurements, the recent proposed Random Finite Set
(RFS) [15–20] avoids explicit measurement-to-track associations This section briefly reviews the formulae of multi-Bernoulli
and computational load becomes more tractable. In the RFS family, RFS and labeled multi-Bernoulli RFS for multi-target tracking. An
the Cardinality Balanced Multi-target Multi-Bernoulli (CBMeMBer) introduction of the multi-target Bayes filter is also presented.
filter propagates the posterior multi-target probability density
by propagating the hypothesized tracks that are parameterized
by the target existence probability and the probability density of 2.1. Multi-target Bayes Filter
the track state [19]. The CBMeMBer filter not only utilizes the
target existence probability for false track discrimination, but also The Multi-target Bayes filter extends the goal of recursive
estimates the multi-target state within tractable computation and state estimation from a single target state xk ∈ X to a multi-target
appears to be an attractive candidate for tracking bearings-only state Xk ∈ X at each scan k, where Xk is composed of a finite
multiple-target with the GMM method. However, the CBMeMBer set of states. The multi-target state Xk distributes according to a
filter operates under the precondition that targets are indistin- multi-target density π k (Xk |Zk ), where Z k = {Zk , Z k−1 } is a sequence
guishable, whereas distinguishing target identities is important of measurement sets up to scan k, and Zk ∈ Z is the multi-
in real tracking systems. By introducing the notion of labeled target observation. The multi-target state Xk and the multi-target
RFS, the δ -generalized Labeled Multi-Bernoulli (δ -GLMB) filter observation Zk are modeled as random finite sets.
[21] enables to provide the target trajectory estimates and their The multi-target density π k (Xk |Zk ) predicts to the next scan k +
unique labels that targets become distinguishable. An efficient im- 1 based on the multi-target Chapman–Kolmogorov formula that
plementation of the δ -GLMB filter called Labeled Multi-Bernoulli the predicted multi-target density πk+1 (Xk+1 |Z k ) is given by [26]
(LMB) filter [22], which outputs target tracks instead of hypotheses 
and achieves better result, has also been proposed. The LMB filter, πk+1|k (Xk+1 |Z k ) = fk+1|k (Xk+1 |X )πk (X |Z k )δ X, (1)
 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44 35

where fk+1|k (Xk+1 |X ) is the multi-target Markov density and the

integral is a set integral defined by [26]
 ∞
 
1
f ( X )δ X = f ( {xi , . . . , xi } )d ( x1 , . . . , xi ). (2)
i! X i
i=0

The multi-target likelihood function is denoted as gk (Zk |Xk ). When

the sensor receives the target observations at scan k + 1, the pos-
terior multi-target density πk+1 (Xk+1 |Z k+1 ) is obtained by [26]

g (Zk+1 |Xk+1 )πk+1|k (Xk+1 |Z k )

πk+1 (Xk+1 |Z k+1 ) = k+1 . (3)
gk+1 (Zk+1 |X )πk+1|k (X |Z k )δ X
Collectively, Eqs. (1) and (3) form one recursion in the multi-target
Bayes filter.

2.2. Multi-Bernoulli RFS

A Bernoulli RFS can either have a probability r ∈ (0, 1) of

containing one element with probability density p(x) or be empty
with a probability 1 − r [27]. Hence, the probability density Fig. 1. Gaussian mixture measurement representation of a bearing measurement.
function (pdf) of a Bernoulli RFS X can be concluded as

1 − r, X = ∅ where
π (X ) = (4)
r p( x ) , X = { x } .   1L (  )r (  )
ω (L ) = ( 1 − r (i ) ) , (11)
A multi-Bernoulli RFS is a union of independent Bernoulli
∈L
1 − r ( )
i∈L
RFSs with cardinality M, i.e., X = ∪M
i=1
X (i ) [19]. The pdf of a
multi-Bernoulli RFS is given by
p(x,  ) = p( ) (x ). (12)
π ({x1 , . . . , xn } ) = {(r (i) , p(i) (x ))}M
i=1 (5)
  In the following sections, the single-target state is notated by
r (i j ) p(i j ) (x j )
n
= π (∅ ) , small letters, e.g. x and x, while multi-target states are notated by
1≤i1 =···in ≤M j=1
1 − r (i j ) capital letters, e.g. X and X. The letters for labeled states, e.g. x
M ( j ) ).
and X, are bolded to be distinct from the unlabeled states, e.g. x
where π (∅ ) = j=1 (1 − r and X. The state space X, measurement space Z and label space L
are notated by blackboard bold.
2.3. Labeled multi-Bernoulli RFS

For the purpose of distinguishing object identities in multi- 3. Gaussian mixture measurements
target scenarios, each target state x ∈ X is allocated with a track
label  ∈ L. The state vector after labeling becomes x = (x,  ). For bearings-only multi-target tracking algorithms, the mea-
Define L: X × L → L as the projection L((x,  )) = , then L(X ) is surement likelihood function is mostly modeled by or approxi-
a label set of a labeled RFS X. The distinct label indicator mated by one Gaussian distribution in the EKF or the UKF, which
may fail to model the measurement uncertainty area precisely. The
(X ) = δ|X | (|L(X )| ), (6)
GMM method separates the measurement uncertainty into several
where subareas, each is modeled by one Gaussian distribution, then the
 measurement likelihood function is approximated by a Gaussian
1, X = Y
δY (X ) = , (7) mixture.
0, otherwise
As shown in Fig. 1, the bearing measurement uncertainty in
ensures that the labels in L(X ) are distinct. Hence, a labeled surveillance space is bounded by a range interval [Lk, min , Lk, max ]
RFS X is a multi-Bernoulli RFS on X augmented with labels on and the sensor noise standard deviation σ φ . The measurement
L [21]. After tagging each target with a unique label, the labeled likelihood function is either modeled by one Gaussian (the ellipse
multi-Bernoulli RFS density is given by bounded by black dashed line) by the EKF or is approximated
by 6 Gaussian components (the ellipses bounded by black solid
π ({(x1 , 1 ), . . . , (xn , n )} ) = δn (|{1 , . . . , n }| ) line) by GMM. The posterior state estimate is bounded by the
  1α ( ) ( j )r (α ( j )) p(α ( j )) (x j )
n −1 −1
intersection of the likelihood function coverage and the predicted
× ( 1 − r (ζ ) ) (α ( ))
−1
, (8) state pdf coverage (the ellipse bounded by red solid line), and the
ζ ∈ j=1
1−r j

estimation error is determined by the intersection precision. In

where  is the label set, α :  → L is a one-to-one mapping, and bearing-only tracking environments, the predicted target position
1α ( ) (j ) is the inclusion function given by is mostly inaccurate before the observability is satisfied, which
 leads to a large predicted error covariance (a large red ellipse). As a
1, X ⊆ Y consequence, the one Gaussian distribution model in the EKF over-
1Y (X ) = . (9)
0, otherwise laps more area, which distributes (specified as shadowed region)
outside the real measurement uncertainty, so that worse estima-
Sometimes the notation 1Y (X) is used as 1Y (x) when X contains
tion results can be expected [28]. Although the Gaussian mixture
only one element X = {x} [22]. According to Eqs. (6) and (8), the
measurements do not cover the entire measurement uncertainty,
pdf of a labeled RFS X is given by
the likelihood becomes more concentrated and contributes to the
π (X ) = (X )ω (L(X )) pX , (10) estimation results after an appropriate GMM division.
36 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44

where N (·; m, P ) denotes a Gaussian distribution with mean m

and covariance P, H is the measurement matrix, and λk is a
coordinate transformation factor given by [30]
 Lk,max L2k,max − L2k,min
λk = RdR = . (21)
Lk,min 2

4. LMB with Gaussian mixture measurements

The LMB filter is a special case and can be interpreted as

an efficient implementation of the δ -GLMB filter. It provides a
closed-form recursion after approximating the first moment of the
posterior multi-target density. The LMB filter inherits the accuracy
of the δ -GLMB filter without introducing exponential growth of
components. For applications where tracking targets using non-
linear measurements, the standard LMB filter usually models the
measurement likelihood function as one Gaussian distribution.
The proposed LMB-GMM filter approximates the track state pdf
as a Gaussian mixture of mutually exclusive and exhaustive track
components. The target state pdf propagates in the prediction step
Fig. 2. An example of parameterized Gaussian mixture measurement with Ak = 3.
and is updated by the Gaussian mixture measurements.

In GMM, the range interval is divided into Ak subintervals 4.1. Track state
under a geometric principle [29]
A labeled multi-Bernoulli RFS is composed of a union of inde-
Lk,a+1
= ρk , a = 1, ., Ak , (13) pendent labeled Bernoulli RFSs. At scan k, each Bernoulli RFS has
Lk,a an existence probability r and track state pdf p(x, ). Accumulating
where Lk,1 = Lk,min and Lk,Ak +1 = Lk,max , and the measurement information up to scan k, Eq. (12) becomes the
posterior track state pdf p(x, |Zk ) and can be approximated by a
 1
Lk,max Ak Gaussian mixture
ρk = . (14)
Lk,min 
Ck
p(x, |Z k ) = ξk (c ) p(x, |c, Z k ), (22)
Define L̄k,a = 0.5(Lk,a+1 + Lk,a ) and Lk,a = Lk,a+1 − Lk,a as the c=1
mean range and length of range interval a, respectively. Using the
where c is the track component index, Ck is the track component
subinterval as the semi-major axis, each subinterval a corresponds
number, ξ k (c)P{c|Zk } is the track component probability, and p(x,
to segment a and generates one Gaussian component a with mean
|c, Zk ) is the track component state pdf. Each track component
yak , covariance Rak , and weight γka [10], with
state pdf is modeled by a Gaussian distribution such that
ρk2 − 1 p(x, |c, Z k ) = N (x, ; xc , Pc ),
γka = ρk2a−2 , (15) (23)
ρk2Ak − 1
where xc and Pc are the mean and covariance of the track com-
Ak ponent state pdf, respectively. Each ξ k (c) indicates the relative
γka = 1. (16) probability that track component c is true among the other
a=1 track components in the same track. A summation of all track
component probabilities satisfies the condition that
Denote zk, i as the ith element in Zk . For each Gaussian component
a generated by measurement zk, i , the mean yak is given by 
Ck
ξk ( c ) = 1 . (24)
p(x,k
s)
cos(zk,i ) c=1
yak = + L̄k,a , (17)
p(y,k
s) sin(zk,i ) Then, each track state pdf in Eq. (10) becomes a Gaussian mixture
(s ) (s )
of track components.
where px,k and py,k are sensor positions in x and y axes. The
corresponding covariance Rak is given by 4.2. Prediction

(Lk,a /2 ) 2
0
Rak = Tk,i T
Tk,i , (18) The labeled multi-Bernoulli RFS predicts forward through
0 (L̄k,a σφ )2 the multi-target Chapman–Kolmogorov formula, which involves
where Tk, i is the rotation matrix given by Markov shifts of existing targets and superposition of newborn
targets [21,22]. After transformation, the pdf of predicted LMB RFS
cos(zk,i ) −sin(zk,i ) X + at scan k + 1 is given by
Tk,i = . (19)
sin(zk,i ) cos(zk,i )
π+ (X + ) = {(r+(,S) , p(+,S) (x|Z k ))}∈Lk ∪ {(rB() , p(B) (x ))}∈B , (25)
An example of the parameterized Gaussian mixture measurement
which is a union of surviving objects from scan k and newborn
is shown in Fig. 2. The measurement likelihood of zk, i for the
objects from scan k + 1.
target with label  can be approximated by
The first term in (25) refers to the pdf of a survival LMB RFS W

Ak given by
f (Hxk , ; zk,i ) ≈ λk γka N (Hxk ; yak , Rak ), (20)
a=1
π+,S (W ) = (W )ω+,S (L(W ))[ p+,S ]W , (26)
 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44 37

4.3. Update
  1L (  )r+( )
ω+,S (L ) = (1 − r+(i,S) ) ( )
,S
, (27) An LMB RFS is not a conjugate prior with respect to the poste-
i∈L ∈L 1 − r +,S rior multi-target state pdf obtained by the multi-target Bayes filter,
with parameters determined by which can be interpreted that the LMB filter is not closed under
the update procedure [21,22]. In order to obtain a posterior multi-
( )
r+ ,S
= ηS (  ) r (  ) , (28) target state that still follows a labeled multi-Bernoulli distribution,
the updated labeled multi-Bernoulli RFS is approximated by its
first moment of the multi-target posterior density. The LMB filter
p(+,S) (x ) = pS (·,  ) f+ (x|·,  ), p(·, |Z k ) /ηS ( ) update transforms the predicted LMB RFS to the relevant δ -GLMB
RFS with compromising a single hypothesis component. For re-

Ck
ducing computational complexity, the LMB RFS and measurements
= ξk (c ) pS (·,  ) f+ (x|·,  ), p(·, |c, Z k ) /ηS ( ), (29)
are partitioned into N approximately statistically independent
c=1
groups, each group operates the predicted LMB RFS to δ -GLMB
transformation as well as the δ -GLMB update separately in parallel.
ηS ( ) = pS (·,  ), p(·, |Z k )

Ck
4.3.1. Group partition
= ξk (c ) pS (·,  ), p(·, |c, Z k ) , (30) In order to improve the computational efficiency, the predicted
c=1
LMB RFS X + performs measurement selection to filter out the
( ) statistically insignificant measurements that invalidated data asso-
where r+ ,S
is the target existence probability of survival tar-
ciations can be avoided. Details of the measurement selection are
gets, p(+,S) (x ) is the relevant single-target state pdf, pS ( · , ) is
available in [22]. The target measurement has a probability pG of
the state-dependent survival probability, fS (x| · , ) is the single-
falling into the validation region [9]. Then, tracks are partitioned
target Markov transition density, and operator · represents
 into groups based on the criterion that tracks in one group share
v, h = v(x )h(x )dx. The state pdf of the survival target p(+,S) (x ) at least one measurement. The predicted multi-target density is
becomes a Gaussian mixture of predicted track components. Given written in the form of
that the survival probability is state independent pS (·,  ) = pS ,
N
Eq. (29) becomes
π + (X + ) = {(r+() , p(+) )}∈L+(i)

Ck i=1
p(+,S) (x ) = ξk (c ) f+ (x|·,  ), p(·, |c, Z k ) N
c=1 = π+(i) (X +(i) ) , (39)

Ck i=1
= ξk (c ) p(x̄, |c, Z k ), (31) (i )
c=1 where π + (X +(i) ) is the predicted multi-target density of the ith
(i )
group, L+ is a subset of label set L+ , and N is the group number.
where x̄ denotes the predicted target state, and p(x̄, |c, Z k ) =
f+ (x|·,  ), p(·, |c, Z k ) is the predicted state pdf of track compo-
nent c that can be obtained by applying the standard Kalman filter 4.3.2. Representation of predicted LMB RFS
prediction formulae. Similarly, we can get The δ -GLMB form of the predicted density for the ith group of
(i )
LMB RFS X + is given by
( )
r+ ,S
= pS r (  ) . (32)  (i )
π+(i) (X +(i) ) = (X +(i) ) ω+(I+,i ) δI+ (L(X +(i) ))[ p+ ]X + , (40)
The second term in (25) is the pdf of a newborn LMB RFS Y (i )
I+ ∈F (L+ )
with label space B given by
with
πB (Y ) = (Y )ωB (L(Y ))[ pB ]Y , (33)

( )
(I+ )
  1L+(i) ( )r+
(  )+
ω+,i = (1 − r ) , (41)
  1B (  )r (  ) ( )
1 − r+


ωB ( L ) = (1 − rB(i) ) B
, (34)
(i )
∈L +  ∈I+

i∈B ∈L 1 − rB( )

where p+ (x,  ) is defined by Eq. (38) and a Gaussian mixture
( ) of track components. Each track component preserves its in-
pB (x,  ) = pB (x ), (35) tegrity through the transformation. More information about this
transformation is available in [22].
where rB( ) is the initial target existence probability, p(B ) (x ) is the
relevant single-target state pdf at scan k + 1.
Combining Eqs. (25), (26) and (33) yields the pdf of the 4.3.3. δ -GLMB update with Gaussian mixture measurements
predicted LMB RFS given by Applying the δ -GLMB update formulae in [21], for a given prior
δ -GLMB RFS π+(i) (X +(i) ), the posterior δ -GLMB distribution at scan
π+ (X + ) = (X + )ω+ (L(X + ))[ p+ ]X + , (36) k + 1 is determined by
where  (i )
π (i) (X (i) |Z k+1 ) = (X (i) ) ω (I+ ,θ ) (Zk+1 )δI+ (L(X (i) ))[ p(θ ) ]X ,
ω+ (I+ ) = ωB (I+ ∩ B )ω+,S (I+ ∩ L ), (37) (I+ ,θ )∈F (L+(i) )× I+

(42)
p+ (x,  ) = 1L ( ) p+,S (x,  ) + 1B ( ) pB (x,  ), (38) (i )
where F (L+ ) denotes the collection of finite subsets of L+ , (i )
I+
denotes the space of data association θ : I+ → {0, 1, 2, ., |Zk+1 |}, and
and the label space is L+ = B ∪ L. The target birth model as well as
the track initiation details are discussed in the following section. ω (I+ ,θ ) (Zk+1 ) ∝ ω+(I+,i ) [ηZ(θk+1) ]I+ , (43)
38 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44

p+,i (x,  )ψZk+1 (x, ; θ )

p(θ ) (x, |Z k+1 ) = , (44)
ηZ(θk+1) ( )

ηZ(θk+1) ( ) = p+,i (·,  ), ψZk+1 (·, ; θ ) , (45)

pD (x, ) pG g(zk+1,θ ( ) |x, )

κ (zk+1,θ () ) , θ ( ) > 0
ψZk+1 (x, ; θ ) = , (46)
qD,G (x,  ), θ ( ) = 0

where g(zk+1,θ () |x,  ) is the single-target measurement likelihood

function, pD (x, ) is the state-dependent target detection proba-
bility, qD,G (x,  ) = 1 − pD (x,  ) pG indicates the target miss-detection
probability, and κ ( · ) is the Poisson clutter intensity.
In the δ -GLMB update, one bearing measurement is divided
into a group of measurement components as defined by Eqs. (13)–
(16). The likelihood of measurement zk+1,θ ( ) conditioned on track
component c is given by
Ak+1

g(zk+1,θ ( ) |x, c,  ) ≈ λk+1 γka+1 N (yak+1,θ ; Hxc , Ska,c+1,θ ), (47)
a=1

Ska,c
+1,θ
= H Pc H T + Rak+1,θ . (48)

The measurement likelihood g(zk+1,θ ( ) |x,  ) in (46) can be ap-

proximated by


Ck
g(zk+1,θ ( ) |x,  ) ≈ ξk (c )g(zk+1,θ () |x, c,  ) (49)
c=1

Since p+,i (x,  ) is a Gaussian mixture of track components as Fig. 3. Flowchart of LMB with Gaussian mixture measurements.
defined by Eqs. (31) and (38), then one track component c
updates with one measurement component a and generates a
new track component (c, a) with posterior track component pdf 1 
p(,i ) = (,i ) ω (I+ ,θ ) (Zk+1 )1I+ ( ) p(θ ) (x, |Z k+1 ). (55)
p(θ ) (x, |c, a, Z k+1 ). The updated pdf can be obtained by applying r (i )
(I+ ,θ )∈F (L+ )× I+
the standard Kalman filter update formulae. Then the posterior
single-target pdf in (44) is approximated by Remarkably, regarding the δ -GLMB update is combinatorial,
Ck Ak+1 the number of hypotheses in each group grows exponentially
 
p(θ ) (x, |Z k+1 ) ≈ ξk+1 (c, a ) p(θ ) (x, |c, a, Z k+1 ) (50) along with the number of tracks and measurements involved. The
c=1 a=1 posterior multi-target density is usually truncated by performing
the m-Best algorithm [31], which only the m most significant hy-
where ξk+1 (c, a ) is the corresponding probability of track compo-
potheses are enumerated to update the predicted density, instead
nent (c, a)
of enumerating all possible hypotheses [21]. Therefore, one pos-
ξk+1 (c, a )  P (θ ) (c, a|Z k+1 ) (51) sible measurement-to-track association θ () can occur repeatedly
pD (x,  ) pG among m hypotheses. In the standard LMB filter, the updated
= ξk ( c ) λ γ a N (yak+1,θ ; Hxc , Rak+1,θ ). tracks by the same association θ () but from different hypotheses
κ (zk+1,θ () ) k+1 k+1
can be handled by track merging. However, due to the internal
The track component number at time k + 1 is given by coupling between track components that this process cannot be
Ck+1 = Ck Ak+1 . (52) simply applied to the LMB-GMM filter. Proper track component
management is required and is discussed in Section 5.1.
4.3.4. LMB update with Gaussian mixture measurements
Since LMB is a special case of δ -GLMB, the updated δ -GLMB 5. Implementations
RFS can be transformed back into a LMB RFS presentation. An
approximation of the updated δ -GLMB RFS, which is presented in In this section, we describe the track component management
a LMB RFS form, is given by that is inevitable and critical for the δ -GLMB update and δ -GLMB
N
to LMB transformation. The track initiation for the newborn target
π (X |Z k+1 ) ≈ π˜ (X |Z k+1 ) = π (i) (X (i) |Z k+1 ) distribution and the track output methods are also presented. The
i=1
flowchart of the LMB-GMM filter is shown in Fig. 3.
N
≈ {(r (,i) , p(,i) )}∈L+(i) (53) 5.1. Track component management
i=1
Without track component control, the number of track com-
with
 ponents increases exponentially in time. Thus, an efficient man-
r (,i ) = ω (I+ ,θ ) (Zk+1 )1I+ ( ), (54) agement method becomes crucial. The component management
(I+ ,θ )∈F (L+(i) )× I+ in LMB-GMM involves two parts: pruning and merging [10].
Component pruning removes the track components with low
 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44 39

Fig. 4. Graph structure for track component management A.

probabilities, and component merging combines the track com- an example of track component management B. In Fig. 5, although
ponents updated by a common measurement component. After track components p(θ1 ) (x, |cˆ1 , a2 , Z k+1 ) and p(θ2 ) (x, |cˆ2 , a2 , Z k+1 )
pruning and merging, the track components are further truncated are from different hypotheses θ 1 and θ 2 , they are both updated by
to a number of Ntop with the largest probabilities. Dues to the measurement component a2 such that track component merging
speciality of the LMB update which needs to transform into should be performed. The formulae for the merged track com-
and update in a δ -GLMB representation, two track component ponent state pdf and track component probability are the same
management methods are required in the δ -GLMB update (track as in track component management A. Distinct from the merging
component management A) and δ -GLMB to LMB transformation in track component management A, the objectives are changed
(track component management B). from merging the components in the same track to merging
the components through different hypotheses. Again, the new
5.1.1. Track component management A track component is re-indexed as c˜ for differentiation while the
The procedure of track component management A is shown in measurement component history is retained.
Fig. 4. Given a survival track p(+,S) (x ) with two track components
and one bearing measurement is divided into 3 measurement 5.2. Track initiation
components. Then each track component “TC #i” (i = 1, 2) up-
dates with each measurement component “MC #j” ( j = 1, 2, 3) The target birth process used in Section 4.2 is described here.
and generates a new track component “TC #ij” with state pdf The RFS based tracking algorithms usually apply the fixed birth
p(θ ) (x, |ci , a j , Z k+1 ). The track components updated by the iden- model in which newborn tracks are initialized at fixed positions
tical measurement components aj are merged into a new track with a constant initial existence probability or weight [15,17,19].
component cˆ with state pdf p(θ ) (x, |cˆ, a j , Z k+1 ) such that In order to cover the entire surveillance region, a large number
Ck+1 of fixed positions are required, which results in a heavy com-
 putational load. This disadvantage becomes more evident when
p(θ ) (x, |cˆ, a j , Z k+1 ) = ξk+1 (ci , a j ) p(θ ) (x, |ci , a j , Z k+1 ) (56)
the target state pdf is modeled by a Gaussian mixture of track
ci =1
components.
and the relevant track component probability is given by In this paper, we use an adaptive birth distribution that ini-
Ck+1 tializes the tracks of newborn targets based on the measurements

ξk+1 (cˆ) = ξk+1 (ci , a j ). (57) from the previous scan [32]. In adaptive birth distribution, not all
ci =1 the measurements are utilized for track initiation such that the
computational load can be significantly reduced. The measurement
All the new track components are re-indexed, e.g., cˆ1 and cˆ2 , that appears in all data association hypotheses is more likely to be
but the measurement component history is retained. Then, track from an existing target that it becomes invalidate for track initi-
component pruning and re-weighting are performed. ation. The distribution of newborn targets at the next scan k + 1
depends on the current scan measurements Zk and is given by
5.1.2. Track component management B
In the δ -GLMB update, the posterior multi-target density ( )
πB,k+1 = {rB,k (z ), p(B,k
)
(x|z )}|Z=1
k|
, (58)
is truncated since only the m most significant hypotheses are
( )
considered. Among the m hypotheses, a measurement zk, 1 can where rB,k (z ) is the initial target existence probability of the track
be associated to track  in different hypotheses, which implies generated by measurement z, and p(B,k
)
(x|z ) is the corresponding
θi ( ) = θ j ( ) = 1. When we gather all the hypotheses in Eq. (55), ( )
the track components from different θ but with identical mea- track state pdf. Details about rB,k (z ) are available in [22]. The
surement component history aj should be merged. Fig. 5 shows initial track state pdf p(B,k
)
(x|z ) is modeled by a Gaussian mixture
40 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44

Fig. 5. Graph structure for track component management B.

of track components such that 5.3. Track output


Ck
p(B,k
)
( x|z ) ≈ ξk (c ) p(B,k
)
(x|c, z ). (59) The track extraction scheme in [22] discriminates the tracks
c=1
which are following the real targets instead of non-existing targets
by applying a valid interval [rval , rout ] of target existence prob-
Specifically, one measurement component a initializes one track ability. Once the posterior target existence probability exceeds
component c with state pdf p(B,k
)
(x|c, z ) = N (x, l; xc , Pc ), and the preset confirmation threshold rout , the relevant track status
transforms from a tentative track into a confirmed track. In the
yak Rak 02×2
xc = , Pc = , (60) subsequent scans, this track maintains a confirmed status as
02×1 02×2 I2×2 · v2max /3
long as its posterior target existence probability is higher than a
where 0m × n and Im × n indicate the m × n zero matrix and identity validation threshold rval .
matrix. The position parts of the mean state xc and covariance For each confirmed track, we output the first moment of the
Pc in each track component are determined by the measurement posterior single-target pdf p(x, |Zk ):

components yak and Rak , and the velocity parts are initialized with
zero and preset max velocity vmax , respectively. The initial track (x,  )out put = yp(y, |Z k )dy (62)
component probability is given by 
= ξk (c, a )x(c,a) ,
ξk (c ) = γka , (61) (c,a )

and the number of track components is equivalent to the number where x(c, a) is the state estimate of track component pdf p(x, |c,
of measurement components, i.e., Ck = Ak . The pseudo-code for a, Zk ).
the track initiation is shown in Algorithm 1.
This paper adopts the one-point initiation method that requires 6. Computational analysis
only the measurements from previous scan. An alternative track
initiation method called two-point initiation can be founded in Due to the presence of clutter measurements, the occurrences
[33]. More details about the track initiation based on Gaussian of target miss-detections, and the data associations between mea-
mixture measurements are available in [10,30]. surements and tracks, the computational complexity of LMB-GMM
filter cannot be presented directly in an analytical form. In this
Algorithm 1 Pseudo-code for track initiation. section, the computational complexities of EKF, UKF, GMM and
1: One measurement z initializes one track with state pdf SMC are presented at the first step. The computational analyses
p(B,k
)
( x|z ). under LMB framework are discussed afterwards. Specially, the
2: for Each measurement z do particle filter version of SMC implementation for bearings-only
3: Apply the GMM method to obtain a set of measurement tracking is also described.
Ak The SMC method propagates a set of weighted particles
components ∪a=1 {yak , Rak }.
that approximate the multi-target posterior density. The particle
4: One measurement component a initializes one track compo-
propagations can be implemented by a sequential importance
nent c with state pdf p(B,k )
(x|c, z ).
resampling (SIR) particle filter, a generic particle filter and a regu-
5: for a=1:Ak do
larized particle filter (RPF). For tracking targets with bearings-only
6: Track state pdf p(B,k )
(x|c, z ) = N (x, l; xc , Pc ). measurements, the SIR particle filter usually cannot give satisfac-
7: Initialize the mean xc with yak . tory results as the resampling procedure introduces the problem
8: Initialize the covariance Pc with Rak and vmax . of loss of diversity among particles that most particles occupy
9: Initialize the track component probability ξk (c ) = γka . the same point in the state space [8]. The RPF that improves the
10: end for sample diversity becomes an appropriate alternative [34] and is
11: The track state pdf is obtained by p(B,k )
( x|z ) = adopted for performance comparison in this paper.
Ck ( )
c=1 k
ξ (c ) pB,k (x|c, z ). The computational complexity of the filters (such as the EKF,
( ) etc.) is analyzed by giving the number of floating-point operations
12: Initialize target existence probability rB,k ( z ).
(flops) in this section. A flop refers to an operation which includes
13: end for
addition/subtraction, multiplication/division between two floating-
point numbers. The equivalent flop (EF) complexity is proposed in
 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44 41

[35] to reflect the flop complexity in matrix and vector operations.

More descriptions for the concepts of flops and EF complexity are
available in [35]. Given the condition that one track is updated by
one measurement, the EF complexity of the EKF is given by

CEKF (n ) = 8n2 + 2n + 2 + d1 + d2 + d3 , (63)

where n is the state dimension, d1 is for the EF complexity of

linearization of the predicted track state, d2 for the calculation of
Gaussian likelihood for a scalar bearing measurement, d3 for the
calculation of the Jacobian matrix. The EF complexity of the UKF
is given by

n3
CUKF (n ) = + 16n2 + 20n + 11 + (2n + 1 )d1 + (2n + 1 )d2 . (64)
3
In the GMM method, each track component updates with each
measurement component as that the corresponding EF complexity
becomes

CGMM (n, Ck , Ak ) = (19n2 + 19n + 8 + d4 + d5 )Ck Ak , (65)

Fig. 6. The target-sensor geometry.
where d4 is the EF complexity for the calculation of Gaussian
likelihood for a 2 × 1 position vector measurement, d5 for the cal-
culation of posterior track component probability ξ k (c, a), Ck is the 7. Simulations
track component number, Ak is the measurement component num-
ber. Compared to the standard particle filter, the RPF has a rela- The simulation tests 100 Monte Carlo trials, and each trial
tively higher computational complexity which can be accounted to consists of 200 scans. The performances are evaluated by Optimal
the regularization step for the calculation of empirical covariance Sub-pattern Assignment (OSPA) [38] distance and estimated target
and the Npt additional generations from the Epanechnikov or the number. The simulated scenario is illustrated in Fig. 6, where
Gaussian kernel [36]. The EF complexity of the RPF is given by multiple targets are observed by a moving sensor platform. The
target state vector at scan k is xk = [ px,k , py,k , p˙ x,k , p˙ y,k ]T . Under
n3 the additive noise assumption, the dynamic evolution of the target
CRPF (n, Npt ) = (5n2 + 6n + 4 + d2 + d6 + d7 )Npt + + 2n2 + 1, (66)
3 is given by

where Npt is the particle number, d6 for resampling procedure, xk = Fk−1 xk−1 + k−1 , (67)
d7 for the random number complexity. A comparison from
where Fk−1 denotes the transition matrix and k−1 is white
Eqs. (63)–(66) reveals that the EKF requires the least flops. The Gaussian process noise with zero mean and covariance Qk−1
EF complexity of the UKF is one-order higher than that of the
EKF (caused by the Cholesky decomposition), but it does not 1 4 1 3
1 Ts T T
involve in the calculation of the Jacobian matrix such that it is Fk−1 = I2×2  , Qk−1 = σν2 · I2×2  4 s
1 3
2 s ,
0 1 T
2 s
Ts2
only slightly heavier than the EKF in practice. The EF complexity
of the GMM method increases linearly with the number of divided (68)
Gaussian components Ak when Ck is limited by the maximum track
where  is the Kronecker product symbol. The sampling in-
component number Ntop [30]. Although both the EF complexity of
terval is Ts = 10 s, and the standard deviation of the process
the GMM method or the RPF grows proportionally with respect
noise is σν = 0.005 m/s2 . The survival probability of target  is
to the track component number or particle number, the RPF in
pS (x,  ) = pS = 0.98.
practice requires a large number of particles to obtain desirable
Each target is detected by the sensor platform with a probabil-
performances that results in a heavier EF complexity.
ity pD (x,  ) = pD = 0.98. The measurement equation is given by
The filters discussed above operate in each hypothesis under
the δ -GLMB framework. LMB is a collection over all hypotheses zk = h(xk(t ) , xk(s ) ) + νk , (69)
generated by δ -GLMB under different track labels. Consequently
the computational complexity of LMB in the update step is  
(t ) (s )
p(y,k
t)
− p(y,k
s)
determined by two factors: (1)the number of the hypotheses h(xk , xk ) = tan −1
, (70)
generated by the m-Best algorithm; (2)the computational com- p(x,k
t)
− p(x,k
s)

plexity in each hypothesis, such as the filter update. Given a

predicted hypothesis with weight ω(h), the m-Best algorithm where ν k is the sensor noise with zero mean and standard devi-
generates m(h ) = ω (h )Jmax  hypotheses, where Jmax is a preset ation σφ = 1◦ ; xk(t ) and xk(s ) are the target state vector and sensor
value [37]. The latter factor is decided by the applied filter (such platform state vector, respectively. The clutter measurements are
as the EKF, etc.) or method which has been described in the uniformly distributed over the bearing interval [0, 360°]. The
last paragraph. Note that the predicted hypothesis weight ω(h) is clutter measurement number follows a Poisson distribution with
implicitly affected by the measurement likelihood g(zk+1,θ ( ) |x,  ), mean 10. The received bearing measurements in one trial are
then numerical analyses for the number of generated hypotheses shown in Fig. 7. The sensor maneuvers twice to satisfy the target
become unavailable. As a result, the computational complexity of observability around time 500 s and 1300 s, respectively.
the LMB-GMM filter as well as the other LMB associated filters The SMC implementation of the LMB filter is realized by the
cannot be presented directly in analytical forms. Alternatively, RPF in this paper for performance comparisons. The EKF, the UKF
the computational comparison can be recognized intuitively by and the RPF are initialized with mean range L̄ = 12, 0 0 0 m and
comparing the execution time. standard deviation σL = 40 0 0 m [4]. A set of N pt = 10 0 0 0 weighted
42 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44

Fig. 7. Bearing measurements in one Monte Carlo run. Fig. 9. Estimated target number comparison.

True
2 Mean
(a) LMB-EKF
StdDev
0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
6

True
Estimated Target Number

2 Mean
(b) LMB-UKF
StdDev
0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
6

True
2 Mean
(c) LMB-GMM
StdDev
0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
6
Time (sec)

Fig. 8. OSPA distance comparison. 4

True
2
(d) LMB-SMC Mean
StdDev
particles is drawn to approximate the initial density of the RPF 0
[4]. The range interval for LMB-GMM is [40 0 m, 20 0 0 0 m] with 0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time (sec)
Ak = 8. The maximum track component number Ntop in each fil-
ter is set to 32. The maximum velocity for track initiation is set to Fig. 10. Cardinality statistics for (a) LMB-EKF, (b) LMB-UKF, (c) LMB-GMM, (d)LMB-
vmax = 8 m/s. The LMB with EKF (LMB-EKF), UKF (LMB-UKF) and SMC.

RPF (LMB-SMC) are also simulated for performance comparisons.

The OSPA distance (with cutoff = 40 0 0 m, order = 2) in Fig. 8 il-
lustrates different performances. The OSPA results can be analyzed inaccuracy leads to a large OSPA error. Similarly, the increasing
by three intervals: time 0 s to 60 0 s; time 60 0 s to 140 0 s; time OSPA at time 1800 is caused by a target death event. The results
1400 s to 2000 s. In the first interval, the targets are unobservable of performance comparison of estimated target number in mean
and all the filters experience large OSPA distance error. In the sec- and standard deviation are depicted in Figs. 9 and 10, respectively.
ond interval, the sensor maneuver satisfies the observability condi- The results show that though the LMB-SMC delivers the best per-
tion such that the localization error starts to decrease. The observ- formances in OSPA, the differences in the estimated target number
ability condition is further improved in the last interval, so that accuracy between the LMB-SMC filter and LMB-GMM filter are
the OSPA error continues decreasing. Among all the intervals, the marginal. Conclusively, the GMM method delivers good sensitivities
LMB-GMM filter have similar estimation errors compared to the in recognizing newborn targets, which contributes to the rapid es-
LMB-SMC filter and shows significant improvements compared to timation of true target number with high accuracy (small standard
the LMB-EKF and the LMB-UFK. deviation). This is due to the fact that the LMB filter utilizes target
In order to validate the robustness of the proposed method for existence probability as a track quality measure, which is a func-
handling a time-varying number of targets, both the target birth tion of measurement likelihood. The GMM method concentrates
and death events are considered in the simulation. New targets on the measurement likelihood approximation that the tracks
arise at times 100 and 200, and the estimated target number following real targets can be confirmed rapidly and correctly.
 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44 43

Table 1 remain the same. The simulation results in Fig. 12 show that the
Execution time for one sampling interval (Ts = 10 s).
LMB-GMM filter again delivers a satisfactory tracking performance.
Method LMB-EKF LMB-UKF LMB-GMM LMB-SMC

Time (sec) 0.146 0.174 0.366 6.738 8. Conclusions

This paper presents the LMB-GMM filter, which approximates

the measurement likelihood and the track state pdf by a Gaus-
sian mixture, for tracking multiple targets using bearings-only
measurements in cluttered environments. The LMB-GMM filter
combines and benefits from the strengths of the Gaussian mixture
measurement method and the labeled multi-Bernoulli filter. The
standard LMB filter for bearings-only target tracking suffers from
either the estimation inaccuracy (the LMB-EKF and the LMB-
UKF) or heavy computational burden (the LMB-SMC filter). The
LMB-GMM filter exploits the advantage of the GMM method to
model the bearing measurement uncertainty more precisely that
satisfactory estimation results can be achieved without exhibiting
the exponential growth in track component number. Although
LMB-SMC provides slightly better tracking performances, the huge
computational load involved with particle filters leads to the
constraint for real applications. However, similar performances
can be achieved by the LMB-GMM filter within considerably
Fig. 11. Average OSPA under different target detection probabilities. reduced computation time, which exhibits more prominences in
practical installations. Further, experiments on target detection
probability and average clutter number validate the robustness of
the proposed filter under different environmental conditions.

Acknowledgments

This work was conducted at High-speed Vehicle Research Cen-

ter of KAIST with the support of Defense Acquisition Program Ad-
ministration (DAPA) and Agency for Defense Development (ADD).

References

[1] Y. Bar-Shalom, X.-R. Li, Estimation and Tracking- Principles, Techniques, and
Software, Norwood, MA: Artech House, Inc, 1993.
[2] S.C. Nardone, V.J. Aidala, Observability criteria for bearings-only target motion
analysis, IEEE Trans. Aerosp. Electron. Syst. (2) (1981) 162–166.
[3] A. Farina, Target tracking with bearings–only measurements, Signal Process. 78
(1) (1999) 61–78.
[4] B. Ristic, S. Arulampalam, N.J. Gordon, Beyond the Kalman Filter: Particle Fil-
ters for Tracking Applications, Artech House, 2004.
Fig. 12. Average OSPA under different average clutter numbers. [5] T.L. Song, Observability of target tracking with bearings-only measurements,
IEEE Trans. Aerosp. Electron. Syst. 32 (4) (1996) 1468–1472.
[6] Y. Bar-Shalom, X.R. Li, T. Kirubarajan, Estimation with Applications to Tracking
and Navigation, John Wiley & Sons, Inc., 2001.
The simulations are conducted on a Windows 7 platform (Intel
[7] S. Julier, J. Uhlmann, H.F. Durrant-Whyte, A new method for the nonlinear
i7-6700 CPU, 16.0GB RAM) and run with the MATLAB program. transformation of means and covariances in filters and estimators, IEEE Trans.
The execution time comparison in Table 1 shows that the LMB- Automat. Contr. 45 (3) (20 0 0) 477–482.
SMC filter requires remarkably more execution time compared [8] M.S. Arulampalam, S. Maskell, N. Gordon, T. Clapp, A tutorial on particle filters
for online nonlinear/non-Gaussian Bayesian tracking, IEEE Trans. Signal Pro-
to the other filters, which makes it inefficient for practical ap- cess. 50 (2) (2002) 174–188.
plications. In scenarios where the sampling interval is not long [9] S. Blackman, R. Popoli, Design and Analysis of Modern Tracking Systems,
enough, for instance Ts = 5 s, the LMB-SMC filter cannot operate Artech House, 1999.
[10] D. Musicki, Bearings only single-sensor target tracking using gaussian mix-
in real-time that real installations become unrealistic. However, tures, Automatica 45 (9) (2009) 2088–2092.
similar performances for LMB-SMC can be achieved by LMB-GMM [11] S. Blackman, Multiple hypothesis tracking for multiple target tracking, IEEE
with significantly reduced computation time. Aerosp. Electron. Syst. Mag. 19 (1) (2004) 5–18.
[12] D. Musicki, R. Evans, Joint integrated probabilistic data association: JIPDA, IEEE
The experiments for target detection probability and average Trans. Aerosp. Electron. Syst. 40 (3) (2004) 1093–1099.
clutter number tests are conducted to investigate the tracking [13] Y.F. Shi, T.L. Song, J.H. Lee, Multi-target tracking in clutter using a high pulse
performance of the proposed method in various environmental repetition frequency radar, IET Radar, Sonar Navigat. 9 (8) (2015) 1047–1054.
[14] J.E. Hopcroft, R. Motwani, J.D. Ullman, Introduction to Automata Theory, Lan-
conditions. In the target detection probability experiment, the
guages, and Computation (Chapter 10), Addison Wesley, 2007.
detection probability is changed while the other scenario settings [15] B.N. Vo, W.K. Ma, The Gaussian mixture probability hypothesis density filter,
are the same as described above. We calculate the average OSPA IEEE Trans. Signal Process. 54 (11) (2006) 4091–4104.
[16] H. Zhang, H. Ge, J. Yang, Y. Yuan, A GM-PHD algorithm for multiple target
from time 1600 s to 20 0 0 s (the observability condition has been
tracking based on false alarm detection with irregular window, Signal Process.
fully satisfied in this interval) and use it as performance evaluation 120 (2016) 537–552.
metric. Fig. 11 shows that the LMB-GMM filter outperforms the [17] B.T. Vo, B.N. Vo, A. Cantoni, Analytic implementations of the cardinalized
LMB-EKF and LMB-UKF while has similar performances compared probability hypothesis density filter, IEEE Trans. Signal Process. 55 (7) (2007)
3553–3567.
to the LMB-SMC at all the detection probabilities. Similarly, the [18] H. Zhang, Z. Jing, S. Hu, Gaussian mixture CPHD filter with gating technique,
average clutter number is varied while the other scenario settings Signal Process. 89 (8) (2009) 1521–1530.
44 Y. Xie, T.L. Song / Signal Processing 151 (2018) 32–44

[19] B.T. Vo, B.N. Vo, A. Cantoni, The cardinality balanced multi-target multi- [30] D. Musicki, T.L. Song, W.C. Kim, D. Nesic, Non-linear automatic target track-
-bernoulli filter and its implementations, IEEE Trans. Signal Process. 57 (2) ing in clutter using dynamic Gaussian mixture, IET Radar, Sonar Navigat. 6 (9)
(2009) 409–423. (2012) 937–944.
[20] J.-L. Yang, H.-W. Ge, An improved multi-target tracking algorithm based on CB- [31] K.G. Murty, Letter to the editoran algorithm for ranking all the assignments in
MeMBer filter and variational Bayesian approximation, Signal Process. 93 (9) order of increasing cost, Oper. Res. 16 (3) (1968) 682–687.
(2013) 2510–2515. [32] S. Reuter, D. Meissner, B. Wilking, K. Dietmayer, Cardinality balanced mul-
[21] B.T. Vo, B.N. Vo, Labeled random finite sets and multi-object conjugate priors, ti-target multi-Bernoulli filtering using adaptive birth distributions, in: Pro-
IEEE Trans. Signal Process. 61 (13) (2013) 3460–3475. ceedings of the 16th International Conference on Information Fusion, 2013,
[22] S. Reuter, B.T. Vo, B.N. Vo, K. Dietmayer, The labeled multi-Bernoulli filter, IEEE pp. 1608–1615.
Trans. Signal Process. 62 (12) (2014) 3246–3260. [33] S. Challa, M. Morelande, D.M. R. Evans, Fundamentals of Object Tracking, Cam-
[23] H. Deusch, S. Reuter, K. Dietmayer, The labeled multi-Bernoulli SLAM filter, bridge University Press, 2011.
IEEE Signal Process Lett 22 (10) (2015) 1561–1565. [34] C. Musso, N. Oudjane, F. Le Gland, Improving regularised particle filters, in:
[24] A.K. Gostar, R. Hoseinnezhad, T. Rathnayake, X. Wang, A. Bab-Hadiashar, Con- Sequential Monte Carlo methods in practice, Springer, 2001, pp. 247–271.
strained sensor control for labeled multi-bernoulli filter using Cauchy–Schwarz [35] R. Karlsson, T. Schon, F. Gustafsson, Complexity analysis of the marginalized
divergence, IEEE Signal Process. Lett. 24 (9) (2017) 1313–1317. particle filter, IEEE Trans. Signal Process. 53 (11) (2005) 4408–4411.
[25] M. Beard, B.-T. Vo, B.-N. Vo, Bayesian multi-target tracking with merged mea- [36] B.G. Sileshi, C. Ferrer, J. Oliver, Particle filters and resampling techniques: im-
surements using labelled random finite sets., IEEE Trans. Signal Processing 63 portance in computational complexity analysis, in: Design and Architectures
(6) (2015) 1433–1447. for Signal and Image Processing (DASIP), 2013 Conference on, IEEE, 2013,
[26] R.P.S. Mahler, Statistical Multisource-Multitarget Information Fusion, Artech pp. 319–325.
House, Inc., Norwood, MA, USA, 2007. [37] B.-N. Vo, B.-T. Vo, D. Phung, Labeled random finite sets and the Bayes multi–
[27] B. Ristic, B.T. Vo, B.N. Vo, A. Farina, A tutorial on bernoulli filters: theory, target tracking filter, IEEE Trans. Signal Process. 62 (24) (2014) 6554–6567.
implementation and applications, IEEE Trans. Signal Process. 61 (13) (2013) [38] D. Schuhmacher, B.T. Vo, B.N. Vo, A consistent metric for performance
3406–3430. evaluation of multi-object filters, IEEE Trans. Signal Process. 56 (8) (2008)
[28] D. Musicki, R.J. Evans, Measurement Gaussian sum mixture target tracking, in: 3447–3457.
2006 9th International Conference on Information Fusion, 2006, pp. 1–8.
[29] T.R. Kronhamn, Bearings-only target motion analysis based on a multihypothe-
sis Kalman filter and adaptive ownship motion control, IEE Proc. - Radar, Sonar
Navigat. 145 (4) (1998) 247–252.