You are on page 1of 22

applied

sciences
Article
A Sector-Matching Probability Hypothesis Density Filter for
Radar Multiple Target Tracking
Jialin Yang 1, *, Defu Jiang 1, *, Jin Tao 1 , Yiyue Gao 2 , Xingchen Lu 1 , Yan Han 1 and Ming Liu 3

1 The Laboratory of Array and Information Processing, College of Computer and Information,
Hohai University, Nanjing 210098, China
2 College of Energy and Electrical Engineering, Hohai University, Nanjing 210098, China
3 The 28th Research Institute of China Electronics Technology Group Corporation, Nanjing 210007, China
* Correspondence: karin_yang@hhu.edu.cn (J.Y.); surfer_jiangdf0801@163.com (D.J.);
Tel.: +86-025-5809-9136 (D.J.)

Abstract: The development of high-tech, dim, small targets, such as drones and cruise missiles,
brings great challenges to radar multi-target tracking (MTT), making it necessary to extend the
beam dwell time to obtain a high signal-to-noise ratio (SNR). In order to solve the problem of radar
sampling time variation exacerbated by extending the beam dwell time when detecting weak targets,
a sector-matching (SM) PHD filter is proposed, which combines the actual radar system with a
PHD filter and quantifies the relationship between the beam dwell time, the false alarm rate and the
detection probability. The proposed filter divides the scanning area into small sectors to obtain actual
multi-target measurement times and rederives the prediction and update steps based on the actual
sampling time. Furthermore, a state correction step is added before state extraction. Applying the
SM structure to the basic Gaussian mixture PHD (GM-PHD) filter and labeled GM-PHD filter, the
simulation results demonstrate that the proposed structure can improve the accuracy of multi-weak-
target state estimation in the dense clutter and can continuously generate explicit trajectories. The
overall real-time performance of the proposed filter is similar to that of the PHD filter.

Keywords: multi-target tracking; radar; random finite sets; sampling time variety; weak targets

Citation: Yang, J.; Jiang, D.; Tao, J.;


1. Introduction
Gao, Y.; Lu, X.; Han, Y.; Liu, M. A
Sector-Matching Probability
Radar has been widely used in the field of military and civilian applications since
Hypothesis Density Filter for Radar
World War II [1–4], and has been applied to complicated scenarios [5,6]. The development
Multiple Target Tracking. Appl. Sci. of stealth technology, small aircraft, drones, cruise missiles, and other modern high-tech
2023, 13, 2834. https://doi.org/ devices brings weak echo energy, resulting in a low signal-to-noise ratio, which places
10.3390/app13052834 higher demands on radar detection and tracking performance. Multi-target tracking (MTT)
is one of the important functions of radar systems and the estimation of a weak target
Received: 18 January 2023
state and trajectory has become a key, hot issue in the field of MTT. In traditional MTT
Revised: 9 February 2023
methods, joint probabilistic data association (JPDA) [7,8], multiple hypothesis tracking
Accepted: 21 February 2023
(MHT) [9], and probabilistic multiple hypothesis tracking [10–12] have mostly been used. In
Published: 22 February 2023
recent years, random finite set (RFS) [1,13] methods have been widely used in sonar [14,15],
autonomous vehicles, robotics [16], and computer vision [17,18] due to them not using data
association procedures. Probabilistic hypothesis density (PHD) filters [19] have become
Copyright: © 2023 by the authors. a hot research topic due to their low computational complexity by propagating the first
Licensee MDPI, Basel, Switzerland. moment of multi-target posteriors. While the PHD filter provides an estimation of target
This article is an open access article states, it provides no target trajectory information.
distributed under the terms and In order to obtain the trajectory of each target, the concept of labeled RFS was intro-
conditions of the Creative Commons duced in the generalized labeled multiple Bernoulli (GLMB) filter [20,21], which improves
Attribution (CC BY) license (https:// the accuracy of ground multi-target state extraction significantly and is suitable for the esti-
creativecommons.org/licenses/by/ mation of explicit target trajectories. Based on the GLMB filter, a joint GLMB filter [22] was
4.0/).

Appl. Sci. 2023, 13, 2834. https://doi.org/10.3390/app13052834 https://www.mdpi.com/journal/applsci


Appl. Sci. 2023, 13, 2834 2 of 22

proposed which combines prediction and updating into one step and reduces the computa-
tional complexity. The GLMB filter [20–22] has difficulty generating the correct multi-target
trajectories in good real-time performance. The trajectory Poisson multi-Bernoulli filter
and the trajectory Poisson multi-Bernoulli hybrid filter proposed in [23] have better fil-
tering accuracy and real-time performance than the joint GLMB filter. Ref. [24] proposed
message-passing-based multitarget tracking methods which can cope with clutter, missed
detections, and unknown associations between targets and measurements. Based on the
general coordinate ascent variational filtering framework, an online VB-AbNHPP tracker
was proposed in [25], which is superior to other competing methods in terms of implemen-
tation efficiency and tracking accuracy. Ref. [26] provided an algorithm for approximating
the inference in this model using a Markov chain Monte Carlo (MCMC)-based auxiliary
variable particle filter, which significantly reduced the computational cost per iteration of
the Markov chain. Ref. [27] introduced two new methods (the peak-fusion method and
the group-target track-before-detect algorithm) for OTHR tracking, which significantly
improved the ability to follow maneuvers. Ref. [28] brought different schools of the tracking
community together by demonstrating that PGFLs are very precise and succinct models of
the combinatorial probability structures are involved in multitarget tracking. All results of
these filters [23–28] have shortcomings in terms of real-time performance.
How to construct a PHD filter suitable for trajectory generation has become the focus
of MTT in recent years, as PHD filters have lower computational complexities. Based on the
output estimates of a PHD filter, a data association step was performed in [29,30] to generate
multi-target trajectories. In [31,32], labels were added to the state estimation of the PHD
filter. In [33], a trajectory probability hypothesis density (TPHD) filter was proposed which
used the trajectory set as the state for MTT. In [34], a Gaussian mixture trajectory PHD filter
was proposed, which allowed trajectory estimation of surviving targets without adding
labels; however, it has difficulty in shielding against the interference from clutter nearby the
targets. In [35], a labeled Gaussian mixture PHD (LGM-PHD) filter was proposed to shield
against the interference and achieve trajectory maintenance by attaching labels to Gaussian
components and introducing an update matrix to obtain target tracking information.
Sampling time variety due to radar beam scanning has not been taken into account
in PHD filters [36]. In contrast to an optical imaging system, where all measurements
within the monitoring space are simultaneously generated at the moment when photos are
taken, radar obtains measurements when the beam illuminates the target. It takes a long
time for the beam to scan the entire airspace and this causes targets at different locations
to be detected at different times. The sampling time difference cannot be ignored, for it
leads to a mismatch between multi-target state prediction and measurement, resulting in
additional measurement errors, especially when the beam scans large surveillance areas
or searches for weak targets and illuminated targets with a long dwell time to obtain a
better signal-to-noise ratio (SNR) [13,37]. Liu previously proposed a time-labeled RFS
which has been applied to a number of problems, such as the Bayesian framework [13] and
the ET-PHD filter [37]. However, specific measurement time information is generally not
included in radar echoes and Liu’s work considered the detection probability as a constant
and did not consider the variation in the detection probability with beam dwell time in
actual radar operation.
In this paper, a novel sector-matching PHD filter is proposed. In order to make the
sector-matching framework more suitable for radar real-time MTT in a complex scenario,
we combine a real radar system with the LGM-PHD filter [35] and quantify the relationship
between beam dwell time, signal-to-noise ratio (SNR), false alarm rate and detection
probability. To simplify the problem, we consider a radar one-way electro-scan scenario,
in which the target position (in particular, the bearing) and target measurement time
correspond one-to-one. The main contributions of this article are as follows:
• The proposed filter divides the detection space into several small sectors and assumes
that the different sectors are independent of each other;
Appl. Sci. 2023, 13, 2834 3 of 22

• Based on these sectors and measurement states, the actual measurement time can
be obtained and used to split and derive the prediction and update equations of the
PHD filter;
• The quantitative relationship between beam dwell time, false alarm rate, detection
probability and SNR is combined with PHD filters;
• Since we interested in the target state at the end of each scan, a state correction step is
added before extracting the target state;
• We evaluate the effectiveness of the proposed method using the constant velocity (CV)
and constant acceleration (CA) dynamic models, and demonstrate that the proposed
filter can solve the problem of high clutter rate and radar sampling time variety when
detecting weak targets well, and enables the track maintenance function.
The rest of the article is organized as follows. Section 2 presents the technical back-
ground. Section 3 presents the specific design of the improved SM PHD multi-filter using
its Gaussian mixture implementations. Quantitative simulation experiments are described
in Section 4. Section 5 presents our concluding remarks.

2. Background
2.1. GM-PHD Filter
In RFS-based filters, the multi-target states and multi-target measurements at the kth
time step are defined as finite sets:
N
n o
Xk = x1k , x2k , · · ·, xk x,k ∈ F (X), (1)

N
n o
Zk = z1k , z2k , · · ·, zk z,k ∈ F (Z), (2)
N
where x1k , x2k , · · ·, xk x,k represent the states of Nx,k targets, F (X) is the set of multi-target state
N
sets X, z1k , z2k , · · ·, zk z,k represent the measurements of Nz,k targets and F (Z) is the set of
multi-target measurement sets Z. The single target motion model and measurement model
are defined as:

xk = Fk −1 xk −1 + wk −1
, (3)
zk = Hk xk + vk
where xk ∈ Rn and zk ∈ Rm are state vector and measurement vector, respectively, Fk−1
represents the target state transition function, Hk is the measurement equation of targets,
wk−1 and vk are the system process noise and measurement noise, distributed as a Gaussian
with variance Qk−1 and Rk , respectively.
The posterior hypothesis posterior PHD at the k − 1 moment can be expressed in the
form of a Gaussian sum as:
Jk−1  
Dk − 1 | k −1 ( x ) = ∑ ωki −1 N x; mik−1 , Pik−1 , (4)
i =1

where Jk−1 is the total number of components and the mean and variance of the ith Gaussian
component are mik−1 and Pik−1 respectively.
The prediction of PHD can be expressed in a Gaussian mixed form as:

Dk|k−1 (x) = DS,k|k−1 (x) + Dβ,k|k−1 (x) + γk (x)


Jk|k−1   (5)
= ∑ ωki |k−1 N x; mik|k−1 , Pik|k−1 ,
i =1

where DS,k|k−1 (x) is the probability hypothesis density of survival targets, Dβ,k|k−1 (x) is the
spawned target probability hypothesis density and γk (x) is the newborn target probability
hypothesis density [19].
Appl. Sci. 2023, 13, 2834 4 of 22

The predicted intensity of the survival target, DS,k|k−1 (x), can be derived from the
target survival rate as well as the prediction equation:

Jk−1  
DS,k|k−1 (x) = pS,k ∑ ωki −1 N x; miS,k|k−1 , PiS,k|k−1 , (6)
i =1

miS,k|k−1 = Fk−1 mik−1 , (7)

PiS,k|k−1 = Fk−1 Pik−1 Fk−1 0 + Qk−1 , (8)

where pS,k denotes the target survival probability and ωki |k−1 = pS,k ωki −1 .The spawned
targets at time k are generated from surviving targets, so that Dβ,k|k−1 (x) can be expressed as:
Jk−1 Jβ,k  
∑ ∑ ωki −1 ωβ,k N
j i.j i.j
Dβ,k|k−1 (x) = x ; mβ,k|k−1 , Pβ,k|k−1 , (9)
i =1 j =1

i,j j j
mβ,k|k−1 = Fβ,k−1 mik−1 + dβ,k−1 , (10)
i,j j j j
Pβ,k|k−1 = Fβ,k−1 Piβ,k−1 Fβ,k−1 0 + Qβ,k−1 , (11)
j j j j
where ω β,k , Fβ,k−1 , dβ,k−1 , Qβ,k−1 , j = 1, · · · , Jβ,k are all spawned target parameters.
The PHD of the newborn targets can also be expressed in a Gaussian mixed form as:

Jγ,k  
γk ( x ) = ∑ ωγ,k
i
N x; miγ,k , Piγ,k , (12)
i =1

i , mi , Pi , i = 1, · · · , J
where ωγ,k γ,k γ,k γ,k are newborn targets parameters.
If the measurement equation can also satisfy the linear Gaussian condition, the updated
PHD at time k can also be expressed in the form of a Gaussian mixture:

Dk|k (x) = (1 − PD,k ) Dk|k−1 (x)


Jk|k−1   (13)
+ ∑ ωki |k (z) N x; mik|k (z), Pik|k ,
i =1
 
PD,k ωki |k−1 N z; Hk mik|k−1 , Sik
ωki |k (z) = Jk|k−1 j
 , (14)
j j
λk (z) + PD,k ∑ j= 1 ω k | k −1
N z; H m ,
k k | k −1 kS
 
mik|k (z) = mik|k−1 (z) + Kik z − Hk mik|k−1 , (15)
h i
Pik|k = I − Kik Hk Pik|k−1 , (16)

Sik = Rk + Hk Pik|k−1 H0k , (17)


  −1
Kik = Pik|k−1 H0k Hk Pik|k−1 H0k + Rk , (18)

where λk is the clutter or false measurements received by the radar and PD,k is the detec-
tion probability.
In order to reduce the computational effort, the increasing Gaussian components must
be pruned and merged [19]. Then the posterior intensity, Dk|k , is expressed as:
Jk  
Dk | k ( x ) = ∑ ωki (z) N x; mik , Pik , (19)
i =1

where Jk is the number of components of Dk|k .


Appl. Sci. 2023, 13, 2834 5 of 22

2.2. Radar Detection


We generally use the method of increasing the beam dwell time (generally considered
to be the accumulation time) to improve the signal-to-noise ratio and target detection
probability, as the pulse-echo signal power, A2i , received by the radar from an individual
target is very low for dim, small targets.
Consider a pulsed radar with pulse width τ, pulse repetition frequency (PRF) f r and
peak transmit power Pt . N pulses are accumulated in one frame, then the radar equation
can be written as:
Pt G2 λ2 σTi f r τ
SNR = , (20)
(4π )3 R4 kT0 FL
where G is the radar antenna gain, λ is the radar operating wavelength, σ is the radar
reflection cross section (RCS) of the target, R is the distance between the target and radar,
k = 1.38 × 10−23 J/K is the Boltzmann constant, T0 = 290 K, Ti = N · τ is the radar beam
dwell time, F is the receiver noise factor and L is the loss of radar. As the accumulated echo
signal power
A2 = ( N · A i )2 , (21)
it can be noted that the SNR is directly proportional to beam dwell time, Ti , and the
amplitude, A, of the accumulated echo signal is also in direct proportion to Ti .
The radar false alarm probability is defined as the probability that the sampling r of
radar echo signal r (t) exceeds the threshold value UT when only noise is present in r (t).
The probability density function of the output voltage of Gaussian white noise after passing
through the envelope detector obeys the Rayleigh distribution as follows:

r2
 
pfa (r ) = r exp − , r ≥ 0, (22)
2

by integrating the above equation, the false alarm probability, Pfa , can be written as:
Z ∞
!
UT2
Pfa = pfa (r )dr = exp − , (23)
UT 2

then, with a given Pfa , the threshold voltage can be found as:
s
1
UT = 2 ln (24)
Pfa

The detection probability, PD , is the probability that a sampling value r of r (t) exceeds
the threshold voltage when the radar echo signal, r (t), contains both noise and signal. After
the Gaussian noise and radar echo signal with amplitude A pass through the envelope
detector, the probability density function of the envelope output by the detector is:
 2
r + A2

pd (r ) = r exp − I0 (rA), r ≥ 0, (25)
2

where I0 (·) is the first class zero-order modified Bessel function.


By integrating Equation (25), the detection probability, PD , can be obtained as follows:
Z ∞ Z ∞  2
r + A2

PD = pd (r )dr = r exp − I0 (rA)dr, (26)
UT UT 2

The relationship between detection probability and false alarm probability is shown
in Figure 1. The area in red represents the false alarm probability, Pfa , the area in purple
represents the detection probability, PD , and the horizontal coordinate of the intersection of
Appl.
Appl.Sci.
Sci.2023,
2023,13,
13,x2834
FOR PEER REVIEW 6 6ofof23
22

of
thethe noise
noise envelope
envelope probability
probability curve
curve and
and thethe envelope
envelope curve
curve of the
of the noise
noise plus
plus signal
signal is
is the
the voltage threshold,
voltage threshold, UT . 𝑈 .

𝑃fa

𝑈
Figure 1. This is the probability density of a filter output for a representative signal-to-noise ratio. The
Figure 1. This is the probability density of a filter output for a representative signal-to-noise ratio.
area under the curve to the right of UT is the probability of detection, PD . Note that while increasing
The area under the curve to the right of 𝑈 is the probability of detection, 𝑃 . Note that while in-
UT decreases Pf a , it also decreases PD .
creasing 𝑈 decreases 𝑃 , it also decreases 𝑃 .
For a single threshold detector, the average time between false alarms for the system
For a single
is calculated threshold detector, the average time between false alarms for the system
as follows:
is calculated as follows:
Ti
tfa = 𝑇 , (27)
𝑡 = fa NTD ,
P
(27)
𝑃 𝑁
where NTD is the number of threshold detectors. The false alarm rate is the frequency of
where 𝑁 occurrence
false alarm per of
is the number unit time, which
threshold is the reciprocal
detectors. of the rate
The false alarm falseisalarm time, thatof
the frequency is:
false alarm occurrence per unit time, which is the reciprocal of the false alarm time, that
1 P N
is: λ= = fa TD , (28)
tfa Ti
1 𝑃 𝑁
𝜆= = , (28)
2.3. Problem Formulation 𝑡 𝑇
The basic GM-PHD filter described in Section 2.1 does not adapt to the radar dynamic
scanning
2.3. Problemmodel and trajectory estimation, especially for detecting weak targets. The
Formulation
method we proposed earlier in Sections 2.2 and 2.3 solves the problem of explicit tracking
The basic GM-PHD filter described in Section 2.1 does not adapt to the radar dynamic
maintenance and is able to solve the problem of misdetection, false state estimation and
scanning model and trajectory estimation, especially for detecting weak targets. The
adverse interference caused by close targets.
method we proposed earlier in Sections 2.2 and 2.3 solves the problem of explicit tracking
In radar systems, the antenna beam width is limited, and it takes a long time to scan
maintenance and is able to solve the problem of misdetection, false state estimation and
the entire airspace, which causes the targets at different locations to be detected at different
adverse interference caused by close targets.
times during a scan, this leads to a variety of sampling times. However, in a basic PHD filter,
In radar systems, the antenna beam width is limited, and it takes a long time to scan
targets are considered to be sampled at the same time, and the measurements generated by
the entire airspace, which causes the targets at different locations to be detected at differ-
different targets are used to update the prediction prior to the same moment, which leads
ent times during a scan, this leads to a variety of sampling times. However, in a basic PHD
to a mismatch between the measurement time and the prediction time.
filter,As
targets
shown areinconsidered
Figure 2, theto be
kthsampled
scan of aatradar
the same time, and
is clockwise, as the measurements
indicated gen-
by θ1 , and the
erated by different targets are used to update the prediction
k + 1th scan is in the counterclockwise direction, indicated by θ2 . prior to the same moment,
whichItleads
can betoseen
a mismatch
that thebetween
samplingthe measurement
times corresponding time to
and the
the prediction time.
measurements of z0k and
0 As shown in Figure 2, the 𝑘th scan of
0 a radar
0 is clockwise,
zk+1 during the kth and k + 1th scans are tk and tk+1 instead of tk and tk+1 . as indicated by 𝜃 , and
the 𝑘If+time
1th scan is in the counterclockwise direction, indicated by 𝜃 .
tk and tk+1 are used in the prediction and update step, the clutter near positions
It can
xk and xk+be seen that the sampling times corresponding to the measurements of 𝑧 and
1 will obtain larger weights and is more likely to be considered as a target
𝑧measurement,
during thewhile𝑘th and 𝑘 + 1th
the actual scans
target are 𝑡 and 𝑡x 0 and
measurements instead of 𝑡 beand
xk0 +1 will 𝑡 . as false
considered
k
If time 𝑡 and 𝑡 are used in the prediction and update step,
alarms, resulting in additional estimation errors and affecting the accuracy of target the clutter near posi-
state
tions 𝑥 and
estimation and 𝑥 trajectory
will obtain larger weights and is more likely to be considered as a tar-
generation.
get measurement, while the actual target measurements 𝑥 and 𝑥 will be considered
as false alarms, resulting in additional estimation errors and affecting the accuracy of tar-
get state estimation and trajectory generation.
as false alarms, resulting in additional estimation errors and affecting the accuracy of ta
In order to solve the above problem and implement a PHD filter-based radar multi-
get state estimation and trajectory generation.
target tracker, a sector-matching PHD filter framework is designed in this paper, which
Appl. Sci. 2023, 13, 2834 In order to solve the above problem and implement a PHD filter-based ofradar mult
involves the measurement time as part of the multi-target state information to7participate 22
target tracker, a sector-matching PHD filter framework is designed in this paper, whic
in prediction, update and state extraction operations.
involves the measurement time as part of the multi-target state information to participat
in prediction, update and state extraction operations.
𝑥
𝑥
𝑥
𝑥 𝑧 𝑥
𝑥
𝑥 𝑥 𝑧
𝑥
𝑥 𝜃
𝑧
𝜃
𝑧 𝜃
Target position 𝜃
Measurement position Measurement times
𝑥 𝑡
Target position Measurement position Measurement times
𝑥𝑥 𝑧 𝑡𝑡
𝑥 𝑥 𝑡 𝑡
𝑧
𝑥 𝑥 𝑡 𝑡
𝑧
𝑥 𝑥 𝑡 𝑡
𝑧
𝑥
Figure 2. Diagram 𝑡 position and the measured position of the
of the relationship between the actual
Figure 2. Diagram of the relationship between the actual position and the measured position of the
target during radar two-dimensional scanning.
target during radar two-dimensional scanning.
Figure 2. Diagram of the relationship between the actual position and the measured position of th
In during
target order toradar
solvetwo-dimensional
the above problem and implement a PHD filter-based radar multi-
scanning.
3.target
Sector-Matching PHD Filter PHD filter framework is designed in this paper, which
tracker, a sector-matching
involves the measurement
A sector-matching timefilter
PHD as part of the multi-target state information to participate
3. Sector-Matching PHD Filter is proposed in this section. We first propose the sector-
in prediction, update and state extraction
matching and sector division issues, and operations.
rederive the prediction and update steps based
A sector-matching PHD filter is proposed in this section. We first propose the secto
on the actual measurement
3.matching
Sector-Matching PHDdivisiontime. Then, a state correction step is given.
Filter
and sector issues, and rederive the prediction and update steps base
on the actual measurement
A sector-matching PHD filtertime. Then, a state
is proposed in thiscorrection step
section. We firstispropose
given. the sector-
3.1. Sector-Matching
matching and Sector
and sector division Division
issues, Issues the prediction and update steps based on
and rederive
the InSector-Matching
themeasurement
3.1.actual process of and radar
time. dynamic
Then,
Sector scanning,
a state
Division Issues thestep
correction direction
is given.that the beam is scanned in
and the sampling time correspond one-to-one [13].
In the process
3.1. Sector-Matching andofSector
radar dynamic
Division scanning, the direction that the beam is scanned i
Issues
For radars in tracking-while-scanning (TWS) mode, the antenna scans at a constant
andInthe sampling time correspond one-to-one [13].
angular the processsoofwe
velocity, radar candynamic
divide the scanning, the direction
surveillance that athe
area into beam isofscanned
number sectorsinequally.
and the For radarstime
sampling in tracking-while-scanning
correspond one-to-one [13]. (TWS) mode, the antenna scans at a constan
As shown in Figure 3, where 𝑧 , 𝑧 , ⋯ 𝑧 are the seven measurements obtained in the 𝑘th
angular velocity,
For radars so we can divide the surveillance
in tracking-while-scanning (TWS) mode, area into a number
the antenna scans at aofconstant
sectors equally
scan, the velocity,
surveillance
so wearea is divided into 𝑛 sectors (𝑛 = 18 in Figure 2) and equally.
the following
angular
As shown in Figure 3,can divide
where 𝑧 the
, 𝑧 surveillance
, ⋯ 𝑧 are the area into
seven ameasurements
number obtained in the 𝑘t
of sectors
assumptions are given. 1 2 , · · · z7 are the seven measurements obtained in the kth
As shown
scan, thein Figure 3, where
surveillance areazkis, zdivided
k k into 𝑛 sectors (𝑛 = 18 in Figure 2) and the followin
scan, the surveillance
assumptions area is divided into n sectors (n = 18 in Figure 2) and the following
are given.
assumptions are given. 100° 90° 80°
110° 70°
120° 90° 60°
100° 80°
110° 70°
130° 50°
120° 60°

140° 𝑧 50°40°
130°
𝑧 𝑧
150° 140° 𝑧 40°
30°
𝑧 𝑧
150° 30°
160° 20°
𝑧 𝑧
160° 20°
170°
𝑧 𝑧 𝑧 10°
𝑧
170° 𝑧 10°
180°
𝑧 0m

2000m

180°
Measurements generated during the 𝑘th scanning 0°
0m 2000m
Figure 3. Radar d
measurements
Measurements generatedduring
generated 𝑘th scanning
in the the 𝑘th scanning
cycle and sector division according to
the sampling time.
Figure
Figure 3. 3. Radar
Radar d
measurements
measurements generated
generated in theinkth
the 𝑘th scanning
scanning cycle
cycle and and
sector sector according
division division according
to t
thesampling
the sampling time.
time.
Appl. Sci. 2023, 13, 2834 8 of 22

Assumption 1. The true state of each target and the generated measurements are independent of
each other [13].

Assumption 2. Each sector is scanned once at most in each scan.

Assumption 3. Any target can only appear in one sector during a scanning cycle.

Assumption 1 is very common in MTT applications. Assumptions 2 and 3 hold in the


case of the TWS operating mode and where the targets do not appear to migrate through
resolution cells (MTRC). In tracking-and-searching (TAS) mode, Assumptions 2 and 3 do
not always hold, which will be developed in later studies.
To simplify the issue, we consider the case of a radar counterclockwise electronic scan
at a uniform angular velocity. The radar scanning period is T and the left and right bound-
aries of the detected airspace are θmax and θmin . The detected airspace is divided equally into
n sectors Sm , m = 1, · · · , n, then the time interval between neighboring sectors is dt = n/T,
the width of the sector is ∆θ = (θmax − θmin )/n and the boundary of sector Sm is denoted as
[θmin + (m − 1)∆θ, θmin + m∆θ ], m = 1, · · · , n. The angle θm of sector Sm is approximated
by the average of the left and right boundaries, i.e., θm = θmin + (m − 1/2)∆θ.
Radar measurements, zik = rki , bki , i = 1, · · · , |Zk |, in the kth scanning cycle, con-


tain the range rki and bearing bki . If bki − θm takes the minimum value when m = mi0 ,

i = 1, · · · , |Zk |, measurement zik falls into the mi 0 th sector, and the measurement time is:
dt
tik = (k − 1) T + mi 0 × dt − , (29)
2
Thus, the multi-target measurement model in Equation (2) is rewritten as:
n   J J
o
Zk = z1k , t1k , · · · , zkt,k , tkt,k , (30)

J
where t1k , · · · , tkt,k is the set of times when different targets are detected in the kth scanning
cycle, correspondingly, then we can rewrite Equation (1) as:
n   J J
  J +1 o
Xk = x1k , t1k , · · · , xkt,k , tkt,k , xkt,k , tk , (31)

where tk is the time at the end of kth scanning cycle and is used to represent the measure-
ment time of the undetected target.

3.2. Initialization and Prediction


Within the kth scanning cycle, the RFSs of the newborn target is expressed in Gaussian
mixture form as shown in Equation (35):
J
   
γt i ( x ) =
γ,k
∑i=γ,k1 ωγ,k
i
N x; miγ,k , tiγ,k , Piγ,k (32)

Each newborn Gaussian component’s sampling time tiγ,k is obtained according to


Equation (32). According to Equation (35), the initial intensity, D0|0 , is described as
n  o Jγ,0
i , Pi
miγ,0 , tiγ,0 , ωγ,0 .
γ,0i =1
The posterior probability density, Dk−1|k−1 (x), at the end of the k − 1th scan can be
n  o Jk−1
j j j
approximated as: mk −1 , t k −1 , ω k −1 , Pk −1 , where tk−1 = (k − 1) T is the timestamp
j =1
of the surviving Gaussian component at time k − 1. The predicted density at the end of kth
scan is:
|Zk |
Dk | k − 1 ( x ) = ∪ Dt i | k − 1 ( x ) (33)
i =1 k
Appl. Sci. 2023, 13, 2834 9 of 22

 J i
j j j j t | k −1
k
Dt i | k − 1 ( x ) = (mt i | k −1 , t k −1 ), ω t i | k −1 , Pt i | k −1 (34)
k k k k j =1

where Jti |k−1 = Jk−1 + Jγ,ti , Jk−1 is the number of components at the k − 1th scanning cycle
k k
and Jγ,ti is the number of newborn components at time tik .
k
Without considering spawned targets, Equations (5)–(8) are rewritten as:
Jti |k−1    
k


j j j j
Dti |k−1 (x) = DS,ti |k−1 (x) + γti (x) = ωi
t k | k −1
N x; m i
t k | k −1
, t k −1 , P i
t k | k −1
(35)
k k k
j =1

Jk−1    

j j j j
DS,ti |k−1 (x)& = pS,ti × ωi N x; m , t k −1 , P (36)
k k t k | k −1 S,tik |k−1 S,tik |k −1
j =1

j j
m & = Ft i | k −1 mk −1 (37)
S,tik |k −1 k

j j
P
S,tik |k −1
& = Ft i | k −1 Pk −1 Ft i | k −1 0 + Qt i | k −1 (38)
k k k

j j
ωi
t k | k −1
= pS,ti ωk−1 (39)
k

j
where the parameters in Fti |k−1 indicate the time interval is tik − tk−1 .
k

3.3. Observation Selection


Considering the real-time performance of the filter, it is most appropriate to remove the
clutter in the low priori probability density region by using the wave gate before updating
the Gaussianh component. Knowing the zero-mean measurement noise covariance matrix,
iT  T
Rk = diag σx2 σy2 , the threshold of the wave gate is d(a) = ak σx σy − [0 0]T k, where


(i) ( j)
k k is the Euclidean distance and a is the confidence coefficient. Then, d(i,j) = kzk − Hk mk|k−1 k,
(i)
i = 1, . . . , |Zk |, j = 1, . . . , Jk|k−1 is computed. The values of zk that satisfies d(i,j) ≤ d(a)
are selected and denoted as efficient observations zik,e f . Then, the efficient observations set
n o|Zk,e f |
Zk,e f = zik,e f , tik,e f is obtained for the kth scanning cycle.
i=1

3.4. Update
Considering the sector-matching issues and combining the PHD filter with actual
radar, the posterior density, Dk|k , at step k can be described as:
|Zk,e f |
Dk | k ( x ) & = ∪ Dt i | t i ( x ) , (40)
i =1 k k

 n o
Dti |ti x& = (1 − PD,k ) Dti |k−1 x +
k k k
∑ DD,ti x; zik,e f , tik,e f
k
(41)
(i )
zk,e f ∈Zk,e f

 n o Jk|k−1 n o  n o 

j j j
DD,ti x; zik,e f , tik,e f = ωi i zik,e f , tik,e f ×N x; m i i zik,e f , tik,e f ,P i , (42)
k tk |tk tk |tk k |tk
i =1
 
j j j
PD,ti ω i N zik ; Hti m i ,S
j
  k t k | k −1 k t | k −1 t ik k
ωi zik = , (43)
tk |tik

Jk|k−1 j j j
λ ti (z) + PD,ti ∑ j= ωi N zik ; Hti m i ,S
k k 1 t k | k −1 k tk |k −1 tik
   
j j j j
mi zik = m i ( z) + K i zik − Hti m i , (44)
tk |tik tk | k −1 t t | k −1
k k k
Appl. Sci. 2023, 13, 2834 10 of 22

 
j j j
Pi i = I − K i Ht i Pi , (45)
tk |tk tk k t k | k −1

j j 0
S i = Rti + Hti Pk|k−1 Hti , (46)
tk k k k
  −1
j j j
Ki = Pi H0 Ht i P i H0 i + Rt i , (47)
tk tk |k −1 tik k t k | k −1 t k k
n     o
j j j
The parameter set mk|k zik,e f , ωk|k zik,e f , tik,e f , Pk|k of the updated Gaussian com-
ponents is obtained, and PD,k , P j and λ j (z) are written, respectively, as:
D,tk t k

Z ∞  2 Z ∞
r + Ak 2

PD,k & = r exp −
pd,k (r )dr = I0 (rAk )dr, (48)
UT UT 2
 2
r + A j2

Z ∞ Z ∞  
tk
P j& = p j (r )dr = r exp −
  I0 rA j dr, (49)
D,tk UT d,tk UT 2 tk

1 P N
λj =
tk
= fa TD , (50)
tfa tint
The undetected targets corresponding to (1 − p D,k ) Dk|k−1 (x) are approximated by the
n o Jk|k−1
parameter set mik|k−1 , (1 − PD,k )ωki |k−1 , tk , Pki |k−1 .
i =1

3.5. Undetected Targets Filtering


In contrast to the detected targets that have a definite sampling time, since we do not
know which sector the missed target corresponds to, it is most appropriate to uniformly
assume that the measurement time of the missed target is at the end of the kth scan. The
parameters of the Gaussian components are:

DU,k (x) = Dk|k−1 (x), (51)

j j
mU,k|k = mU,k|k−1 , (52)
j j
PU,k|k = PU,k|k−1 , (53)
j j
ω k | k = ω k | k −1 , (54)

where U represents the undetected targets.

3.6. Pruning and Merging


Then, the Gaussian components are pruned and merged. This procedure also re-
quires pruning and merging the sector-matching information together. The new Gaussian
component after merging is:
n
∑ ωtik |tik ,
j
e tli |ti =
ω (55)
k k
j =1
n
1
∑ ωtik |tik mktik |tik ,
j j
e lti |ti =
m (56)
k k e l i i j =1
ω
tk |tk
" T #
n  
e lti |ti & = 1

j j j j
P ωi i × Pi i + e lti |ti
m − mt i | t i e lti |ti
m − mt i | t i , (57)
k k e li
ω tk |tk tk |tk k k k k k k k k
tk |tik j=1

and the actual measurement time corresponding to the new Gaussian component after
merging is:
scan, a correction to the posterior is required:
𝒎 | &=𝑭 |
𝒎 |
,
Appl. Sci. 2023, 13, 2834 11 of 22
𝜔 | &=𝜔 ,
|

note that 𝒎 | is only used lfor state n extraction and not for prediction and u
et i i = 1 ∑ t ji i , (58)
ations. Then, 𝒎 | is used totk |tk
carry
n j=out
1 k subsequent
t |tk state extraction operatio
3.7. State Correction
4. Simulations
In the state extraction step, since we aim to obtain the target state at the end of each
scan, aIn this section,
correction we first
to the posterior simulate and analyze the variation in the
is required: false a
detection probability with accumulation time for several specific false alarm
m̂lk|k & = Fk|ti m e lti |ti , (59)
in radar operation. Then, the results acquired using the proposed method
k k k

tracking simulations are presented. ω̂kl |k & = ωWe


j
, apply the proposed method (60)to the P
tik |tik
LGM-PHD filter proposed by Gao in [35], and compare them with the b
note that m̂lk|k is only used for state extraction and not for prediction and updating opera-
filter [19] and the LGM-PHD filter [35]. Simulations are designed using th
tions. Then, m̂lk|k is used to carry out subsequent state extraction operations.
constant velocity (CV) motion model and a constant acceleration (CA) mo
dense
4. scenarios.
Simulations
In this section, we first simulate and analyze the variation in the false alarm rate and
4.1. False
detection Alarm Rate
probability and Detection
with accumulation timeProbability
for several specific false alarm probabilities
in radar operation. Then, the results acquired using the proposed method in multi-target
trackingWith other are
simulations radar parameters
presented. We applykept constant,
the proposed we simulated
method andand
to the PHD filter analyzed
in the false
LGM-PHD alarm
filter ratebyand
proposed Gao detection probability
in [35], and compare them with
with theaccumulation
basic GM-PHD time fo
filter [19] and the LGM-PHD filter [35]. Simulations are designed
cific false alarm probabilities. For radar MTT, a lower false alarm using the multi-target
rate and hi
constant velocity (CV) motion model and a constant acceleration (CA) motion model in
probability
dense scenarios.is desired.
As shown in Figure 4, reducing the constant false alarm detection thre
4.1. False Alarm Rate and Detection Probability
creasing the false alarm probability) can raise the detection probability, but
With other radar parameters kept constant, we simulated and analyzed the variation
antheincrease
in false alarminrate
theandfalse alarm
detection rate. As
probability theaccumulation
with beam dwell timetime increases,
for several specific the de
bility
false increases
alarm and the
probabilities. falseMTT,
For radar alarm rate false
a lower decreases.
alarm rate Therefore, to improve th
and higher detection
probability
of radar is desired.
MTT, a suitable choice is to extend the accumulation time.
As shown in Figure 4, reducing the constant false alarm detection threshold (i.e.,
In the
increasing following
the false simulation
alarm probability) presented
can raise in this
the detection paper,but
probability, theit false alarm pro
also leads
radar
to an increase the𝑃
is setinto false=alarm
0.3,rate.
andAsitthe
can be dwell
beam seentime
from Figure
increases, that 𝜆 prob-
the4detection = 104 and
𝑇 = 2𝑚𝑠, and 𝜆 = 52 and 𝑝 , = 0.96 at 𝑇 = 4𝑚𝑠.
ability increases and the false alarm rate decreases. Therefore, to improve the performance
of radar MTT, a suitable choice is to extend the accumulation time.
Detection Probability
False Alarm Rate

Figure
Figure 4. Variation
4. Variation in falsein false
alarm ratealarm rate and
and detection detection
probability probabilitytime
with accumulation with
withaccumulation
different t
detection probabilities
ent detection for the givenfor
probabilities radar
thein given
the sameradar
scenario.
in the same scenario.
In the following simulation presented in this paper, the false alarm probability of the
4.2. Example
radar is set to Pfa1:=CV
0.3, Motion
and it canModel
be seen from Figure 4 that λ = 104 and p D,k = 0.8 at
Ti = 2 ms, and λ = 52 and p D,k = 0.96 at Ti = 4 ms.
In this scenario, each target moves according to the following Gaussian
Appl. Sci. 2023, 13, 2834 12 of 22

Appl. Sci. 2023, 13, x FOR PEER REVIEW 12 of 23


4.2. Example 1: CV Motion Model
In this scenario, each target moves according to the following Gaussian linear model:

11 T𝑇s ⊗ 𝑰 𝒙 T𝑇s /4⁄4 T ⁄2


𝑇s 3 /2
    2  
𝒙
xk == + 𝜎
2
⊗ Id xk−1 + σw ⊗ 𝑰
⊗ Id (61)
(61)
00 11 ⁄2 T
T𝑇s 3 /2 𝑇s 2

where 𝑰Id is
where is the
the identity
identity matrix
matrix of of dimension
dimension d, 𝑑 , where
where d𝑑 = = 22 in
in this scenario. 𝒙x ==
this scenario.
k
[𝒙 , , 𝒙 , , 𝒙 , , 𝒙 , ]T , [𝒙 , , 𝒙 , T] represents the target position at the end of the 𝑘th scan
[x1,k , x2,k , x3,k , x4,k ] , [x1,k , x3,k ] represents the target position at the end of the kth scan
and [𝒙 , , 𝒙 , ]T is the target velocity. 𝑇 is the sampling period and 𝜎 = 5𝑚/𝑠 .
and [x2,k , x4,k ] is the target velocity. Ts is the sampling period and σw = 5 m/s2 .
The actual target trajectories and the initial states of the targets are given in Figure 5
The actual target trajectories and the initial states of the targets are given in Figure 5
and Table 1, targets are created from the black dot ●, and disappear at the end of their
and Table 1, targets are created from the black dot •, and disappear at the end of their
respective tracks. The multi-target measurement model is given by Equation (60):
respective tracks. The multi-target measurement model is given by Equation (60):
 𝒓 (𝑯 ⊗ 𝑰 )𝒙   𝒓(𝒆 )
𝒛 = r((Hk ⊗ Id )xk ) + r(ek ) (62)
zk = 𝜽 (𝑯 ⊗ 𝑰 )𝒙 + 𝜽(𝒆 ) (62)
θ((Hk ⊗ Id )xk ) θ(ek )
where 𝒓 (𝑯 ⊗ 𝑰 )𝒙 and 𝜽 (𝑯 ⊗ 𝑰 )𝒙 denote the range and bearing of state 𝒙 ,
where r((Hk ⊗
respectively. 𝑯 Id=)x[1 k ) 0]
andisθthe k ⊗ Id )xk ) denote
((Hobservation matrixtheand 𝒆 isand
range bearing of
a Gaussian xk , respec-
statenoise
white with
tively.
covarianceH k 𝑑𝑖𝑎𝑔 ([10,0.1]).
= [ 1 0 ] is the observation matrix and e k is a Gaussian white noise with covariance
diag ([10, 0.1]).

90°
105° 75°

120° 60°

135° 45°

150° 30°

165° 15°

180° 0°
0m 400m 800m 1200m 1600m 2000m

Figure 5. Target trajectories in example 1 (CV motion model), targets are created from the black dot •.
Figure 5. Target trajectories in example 1 (CV motion model), targets are created from the black
dot ●. 1. List of initial target states in example 1 (CV model).
Table

Table 1. List of initial target states in example 1 (CV model). Initial States
Target Index Lifetime (s)
(m,m/s,m,m/s)
Initial States
Target Index
No. 1 Lifetime (s) (1,70) [0; 0; 0; −10]
(𝐦, 𝐦/𝐬, 𝐦, 𝐦/𝐬)
No. 2 (1,100) [400; −10; −600; 10]
No.No.
1 3 (1,70) (1,70) [0; 0;
[−0; −10]
800; 20; −200; −10]
No.No.
2 4 (1,100) (20,100) [400; −10; −7;10]
−600;
[400; −600; −4]
No.No.
3 5 (1,70) (20,100) [−800; 20; [400; −2.5;
−200; −600; 10]
−10]
No. 6 (20,100) [0; 7.5; 0; −5]
No. 4 (20,100) [400; −7; −600; −4]
No. 7 (40,100) [−800; 12; −200; 7]
No.No.
5 8 (20,100) (40,100) [400; −2.5; −600;
[−200; 15;10]
800; −10]
No.No.
6 9 (20,100) (60,100) [0; 7.5; 0; −5]
[−800; 3; −200; 15]
No.No.
7 10 (40,100) (60,100) [−200;
[−800; 12; −200;−3;7]800; −15]
No. 11 (80,100) [0; −40; 0; −30]
No. 8 (40,100) [−200; 15; 800; −10]
No. 12 (80,100) [−200; 30; 800; −10]
No. 9 (60,100) [−800; 3; −200; 15]
No. 10 (60,100) [−200; −3; 800; −15]
As shown
No. 11 in Figures 6 and(80,100)
7, the sector-matching method[0; proposed
−40; 0; −30] in this paper can
achieve No.
radar12MTT when considering
(80,100) radar measurement time30;
[−200; diversity.
800; −10]Applying the
method to the LGM-PHD filter can achieve tracking maintenance even in the presence
of strong clutter interference, dense targets in the surveillance area and long radar scan
As shown in Figures 6 and 7, the sector-matching method proposed in this paper can
periods. In Figure 6b,d, it can be seen that a low detection probability of p D,k = 0.8
achieve radar MTT when considering radar measurement time diversity. Applying the
causes the trajectory to be broken, and the reason why the trajectory can be identified
method to the LGM-PHD filter can achieve tracking maintenance even in the presence of
as the same target’s trajectory instead of a new-born target trajectory after the trajectory
strong clutter interference, dense targets in the surveillance area and long radar scan pe-
is broken is that in the simulation made in this paper, it is assumed that the newborn
riods. In Figure 6b,d, it can be seen that a low detection probability of 𝑝 , = 0.8 causes
the trajectory to be broken, and the reason why the trajectory can be identified as the same
target’s trajectory instead of a new-born target trajectory after the trajectory is broken is
Appl. Sci. 2023, 13, x FOR PEER REVIEW 13 of 23
Appl. Sci. 2023, 13, 2834 13 of 22

that in the simulation made in this paper, it is assumed that the newborn targets can only
appearcan
targets at the
onlyknown
appearnew-born points,
at the known i.e., it ispoints,
new-born assumedi.e.,that
it is no newborn
assumed thattarget will ap-
no newborn
pear at
target theappear
will trajectory breaking
at the points.
trajectory In Figure
breaking points.7b,d, due to 7b,d,
In Figure the extension ofextension
due to the the accumu-
of
the accumulation
lation time, theprobability
time, the detection 𝑝 , = 0.96pand
detection probability D,k = 0.96
the and the
probability probability
of track of track
breakage is
breakage is significantly
significantly reduced. reduced.

90°
105° 75°

120° 60°

135° 45°

150° 30°

165° Estimates 15°


True tracks
Measurements

180° 0°
0m 400m 800m 1200m 1600m 2000m

(a) (b)

(c) (d)
Figure 6. The estimations of the multi-target trajectories and the actual measurements in example 1
Figure 6. The estimations of the multi-target trajectories and the actual measurements in example 1
(CV motion model) for a sample run. (a) The actual measurements, actual trajectories and estimated
(CV motion model) for a sample run. (a) The actual measurements, actual trajectories and estimated
trajectories obtained after filtering by the SM-GM-PHD filter for 𝑇 = 1 s ( 𝑇 = 2 ms, 𝜆 = 104,
trajectories
𝑝 , = 0.8).obtained after measurements,
(b) The actual filtering by theactual
SM-GM-PHD filter
trajectories andfor Ts = 1 trajectories
estimated s (Ti = 2 ms, λ = 104,
obtained after
D,k =
pfiltering0.8). (b) The actual measurements, actual trajectories and estimated trajectories
by the SM-LGM-PHD filter for 𝑇 = 1 s, and the different trajectories are denoted by differ-obtained after
filtering
ent color bydots.
the SM-LGM-PHD
(c) The actual filter for Ts = 1 s,actual
measurements, and the different trajectories
trajectories and estimatedare denoted by different
trajectories obtained
color
afterdots. (c) The
filtering actual
by the measurements,
SM-GM-PHD filteractual trajectories
in x and for 𝑇 = 1trajectories
and estimated
y coordinates s. (d) The actual
obtained after
measure-
ments, by
filtering actual trajectories and
the SM-GM-PHD estimated
filter in x andtrajectories obtained
y coordinates for Ts =after
1 s.filtering by themeasurements,
(d) The actual SM-LGM-PHD
filter in
actual x and y coordinates
trajectories and estimated 𝑇 = 1 s. obtained after filtering by the SM-LGM-PHD filter in x
for trajectories
and y coordinates for Ts = 1 s.

As shown in Figure 8, the computation time (CT) of the MATLAB simulation for
100 sample runs is given and the average CT is shown in Table 2. The basic GM-PHD
filter and the modified SM-GM-PHD filter have relatively smooth computation times.
However, for the LGM-PHD filter, as there is an extra step in which the matrix U is used
to determine trajectories, the computation speed of this process depends on the distance
between the Gaussian components and the weight of the components. If a large number of
measurements are generated in a given scan cycle, this also makes the computation time
higher. Therefore, the computation times of the LGM-PHD and the SM-LGM-PHD filters
vary randomly. The results show that the real-time performance of the proposed method
for both the LGM-PHD filter and the basic GM-PHD filter is similar to that of the basic
PHD filter and the LGM-PHD filter, and much better than that of the joint GLMB filter. In
addition, extending the accumulation time can significantly reduce the calculation time.
Appl.
Appl.Sci.
Sci.2023,
2023,13,
13,x2834
FOR PEER REVIEW 1414ofof23
22

90°
90° 105° 75°
105° 75°

120° 60° 120° 60°

135° 45° 135° 45°

150° 30° 150° 30°

Estimates
True tracks
165° 15° 165° 15°
Estimates Measurements
True tracks
Measurements

180° 0° 180° 0°
0m 400m 800m 1200m 1600m 2000m 0m 400m 800m 1200m 1600m 2000m

(a) (b)

(c) (d)
Figure
Figure7.7.The
Theestimations
estimationsofofthe
themulti-target
multi-targettrajectories
trajectoriesand
andthetheactual
actualmeasurements
measurementsininexample
example11
(CV motion model) for a sample run. (a) The actual measurements, actual trajectories
(CV motion model) for a sample run. (a) The actual measurements, actual trajectories and estimated and estimated
trajectories obtained after filtering by the SM-GM-PHD filter for 𝑇 = 2 s (𝑇 = 4 ms, 𝜆 = 52, 𝑝 , =
trajectories obtained after filtering by the SM-GM-PHD filter for Ts = 2 s (Ti = 4 ms, λ = 52,
0.96). (b) The actual measurements, actual trajectories and estimated trajectories obtained after fil-
p D,k = 0.96). (b) The actual measurements, actual trajectories and estimated trajectories obtained
tering by the SM-LGM-PHD filter for 𝑇 = 2 s, and the different trajectories are denoted by different
afterdots.
color filtering by actual
(c) The the SM-LGM-PHD
measurements, filter
actual Ts = 2 s, and
for trajectories andthe different
estimated trajectories
trajectories are denoted
obtained after
ppl. Sci. 2023, 13, x FOR PEER REVIEW
by different
filtering by color
the dots.
SM-GM-PHD (c) The actual
filter in x measurements,
and y coordinates actual
for 𝑇 = 2 s.
trajectories
(d) and
The estimated
actual trajectories
measurements, 15 o
obtained
actual after filtering
trajectories by the SM-GM-PHD
and estimated filter in
trajectories obtained x and
after y coordinates
filtering for Ts = 2 s.filter
by the SM-LGM-PHD (d) in
The
x
and y coordinates
actual measurements,for 𝑇 actual
= 2 s. trajectories and estimated trajectories obtained after filtering by the
SM-LGM-PHD filter in x and y coordinates for Ts = 2 s.
As shown in Figure 8, the computation time (CT) of the MATLAB simulation for 100
sample runs is given and the average CT is shown 0.14 in Table 2. The basic GM-PHD filter
0.15
and the modified SM-GM-PHD filter have relatively 0.12 smooth computation times. How-
ever, for the LGM-PHD filter, as there is an extra step in which the matrix 𝑼 is used to
determine trajectories, the computation speed 0.1 of this process depends on the distance
SM-LGM-PHD filterbe-
0.1
SM-LGM-PHD filter LGM-PHD filter
tween the Gaussian components and the weight
LGM-PHD filter
0.08 of the components. If a large number of
SM-GM-PHD filter
measurements
SM-GM-PHD filter are generated in a given scan cycle, this also makes the computation
Basic GM-PHD filtertime
Basic GM-PHD filter 0.06 joint-Glmb filter
higher. Therefore,
joint-Glmb filter
the computation times of the LGM-PHD and the SM-LGM-PHD filters
0.05
vary randomly. The results show that the real-time 0.04 performance of the proposed method
for both the LGM-PHD filter and the basic GM-PHD filter is similar to that of the basic
0.02
PHD filter and the LGM-PHD filter, and much better than that of the joint GLMB filter. In
0 0
20 40
addition,
60 80
extending
100
the accumulation time can significantly reduce the calculation time.
20 40 60 80 100
Time/s Time/s
(a) (b)
Figure 8. The
Figure 8. Thecomputation time
computation time of different
of different filters filters for 100runs
for 100 sample sample
at Ts =runs i =𝑇
1 s (Tat = 1λ s= (𝑇
2 ms, 104,= 2 ms,
𝑝 ,==
104, p D,k 0.8)
0.8) s = 2𝑇
(a) Tand
(a) and s (T= 2 4s ms,
i = (𝑇 λ== 452,ms,
p D,k𝜆==0.96)
52,(b)𝑝 in, example
= 0.96)1 (CV
(b) motion
in example
model).1 (CV mo
model).

In Figures 9 and 10, the OSPA distances of the MATLAB simulation for 100 sam
runs is given and the average OSPA distances are given in Table 2. Compared with
basic GM-PHD filter, the average tracking performance of the SM-LGM-PHD filter,
LGM-PHD filter, the SM-GM-PHD filter and the joint-GLMB filter in this scenario is
Appl. Sci. 2023, 13, 2834 15 of 22

Table 2. List of the average CT and average OSPA of each filter in example 1 (CV model).

SM-LGM- LGM- SM-GM- Basic GM- Joint-


PHD Filter PHD Filter PHD Filter PHD Filter GLMB Filter
CT ( Ts = 1 s ) 0.0164 s 0.0180 s 0.0130 s 0.0069 s 0.1125 s
CT ( Ts = 2 s ) 0.0097 s 0.0090 s 0.0078 s 0.0058 s 0.1009 s
OSPA ( Ts = 1 s ) 25.66 m 28.71 m 30.60 m 33.18 m 12.51 m
OSPA ( Ts = 2 s ) 16.62 m 21.61 m 18.20 m 28.17 m 14.80 m

In Figures 9 and 10, the OSPA distances of the MATLAB simulation for 100 sample
runs is given and the average OSPA distances are given in Table 2. Compared with the basic
GM-PHD filter, the average tracking performance of the SM-LGM-PHD filter, the LGM-
PHD filter, the SM-GM-PHD filter and the joint-GLMB filter in this scenario is improved
by 22.66%, 13.47%, 7.78% and 62.29%, respectively, at Ts = 1 s (Ti = 2 ms, λ = 104,
p D,k = 0.8), and 41%, 13.47%, 35.39% and 47.46%, respectively, at Ts = 2 s (Ti = 4 ms,
λ = 52, p D,k = 0.96). Compared with the Ts = 1 s scenario, the performance of the
SM-LGM-PHD filter and the SM-GM-PHD filter has been significantly improved. The
LGM-PHD filter, the basic GM-PHD filter and the joint-GLMB filter are not time-matched,
which is why the performance of these three filters is not that significantly improved, and
the performance of the joint-GLMB filter even decreases. In addition, it can be seen in
Figures 9b and 10b that the values of the OSPA location of the LGM-PHD filter, the basic
GM-PHD filter and the joint-GLMB filter all increase with the accumulation of time. After
80 s, with the appearance of two targets, No.11 and No.12, at a higher speed, the targets’
position estimation errors of the LGM-PHD filter, the basic GM-PHD filter and the joint-
Appl. Sci. 2023, 13, x FOR PEER REVIEWGLMB filter increase at a faster rate, while the SM-LGM-PHD filter and the SM-GM-PHD 16 of 23
filter are not significantly affected. This is mainly caused by the additional measurement
error caused by the sampling time diversity mentioned earlier.
OSPA Dist
OSPA Loc
OSPA Card

Figure
Figure Theaverage
9.9.The averageOSPA
OSPA of of different
differentfilters in in
filters example
example1 (CV motion
1 (CV model)
motion for 100
model) forsample runs runs
100 sample
when T = 1 s (T = 2 ms, =
when 𝑇 = 1 s (𝑇 i= 2 ms, 𝜆 = 104, 𝑝 D,k 104, p = 0.8). (a) Average OSPA distances of different filters.
, = 0.8). (a) Average OSPA distances of different filters. (b)
s λ
(b) OSPA
OSPA locations
locations of different
of different filters.
filters. (c)(c) OSPAcardinalities
OSPA cardinalities of
ofdifferent
differentfilters.
filters.
OSPA Dist
Figure 9. The average OSPA of different filters in example 1 (CV motion model) for 100 sample runs
Appl. Sci. 2023, 13, 2834 when 𝑇 = 1 s (𝑇 = 2 ms, 𝜆 = 104, 𝑝 , = 0.8). (a) Average OSPA distances of different 16 offilters.
22 (b)
OSPA locations of different filters. (c) OSPA cardinalities of different filters.

OSPA Dist
OSPA Loc
OSPA Card

Figure10.
Figure Theaverage
10.The average OSPA
OSPA of ofdifferent
differentfilters in in
filters example 1 (CV
example motion
1 (CV model)
motion for 100for
model) sample
100 sample
runs, when T = 2 s (T = 4 ms, = 52, p = 0.96). (a) Average OSPA distances of
runs, when 𝑇 s = 2 s (𝑇i = 4 ms, 𝜆 = 52,D,k𝑝 , = 0.96). (a) Average OSPA distances of different
λ different filters. fil-
(b) OSPA locations of different filters. (c) OSPA cardinalities of different filters.
ters. (b) OSPA locations of different filters. (c) OSPA cardinalities of different filters.
Comparing the results in Figures 9 and 10, it can be found that the accuracy of multi-
4.3. Example
target 2: CA Motion
state estimation Model with what we expected. The estimation accuracy of the
is consistent
filter
In deteriorates as each
this scenario, the scan period
target increases
moves and also
according deteriorates
to the followingwith the increase
Gaussian linearinmodel:
target velocity.
1 𝑇 𝑇 ⁄2
𝒙 =
4.3. Example 2: CA Motion Model 0 1 𝑇 ⊗𝑰 𝒙 +𝒘 (63)
0 according
In this scenario, each target moves 0 1 to the following Gaussian linear model:

Ts 2 /2
  
1 Ts
xk =  0 1 Ts  ⊗ Id xk−1 + wk (63)
0 0 1

where Id is the identity matrix of dimension d, where d = 2. xk = [x1,k , x2,k , x3,k , x4,k , x5,k , x6,k ] T ,
[x1,k , x4,k ] T represents the target position at the end of the kth scan, [x2,k , x5,k ] T is the target
velocity and [x3,k , x6,k ] T is the target acceleration. Ts is the sampling period. The actual target
trajectories and the initial states of the targets are given in Figure 11 and Table 3, targets are
born from the black dot •, and disappear at the end of their respective tracks. The multi-target
measurement model is given by Equation (62):
   
r((Hk ⊗ Id )xk ) r (ek )
zk = + (64)
θ((Hk ⊗ Id )xk ) θ(ek )

where r((Hk ⊗ Id )xk ) and θ((Hk ⊗ Id )xk ) denote the range and bearing of state xk , respec-
tively. Hk = [1 0 0] is the observation matrix and ek is the Gaussian white noise with
covariance diag ([10, 0.1]).
The radar divides the detection airspace into 500 parts. A detailed comparison of
the performance of the individual filters at Ts = 1 s, Ti = 2 ms, λ = 104, p D,k = 0.8 and
Ts = 2 s, Ti = 4 ms, λ = 52, p D,k = 0.96, and the effect of different sampling times on the
filter performance are given in Figures 12–16.
𝜽 (𝑯 ⊗ 𝑰 )𝒙 𝜽(𝒆 )

where 𝒓 (𝑯 ⊗ 𝑰 )𝒙 and 𝜽 (𝑯 ⊗ 𝑰 )𝒙 denote the range and bearing


respectively. 𝑯 = [1 0 0] is the observation matrix and 𝒆 is the Gaussian
Appl. Sci. 2023, 13, 2834 with covariance 𝑑𝑖𝑎𝑔 ([10,0.1]). 17 of 22

Figure 11. Target trajectories in example 2 (CA motion model), where targets are born from the black
Figure 11. Target trajectories in example 2 (CA motion model), where targets are born
dot •.
black dot ●.
Table 3. List of initial target states in the CA motion model.

Table 3. List of initial target states in the CA motion model.


Initial States
Target Index Lifetime (s)
(m,m/s,m/s2 ,m,m/s,m/s2 )
[0; 0; 0.03; 0; −Initial
37;0,75] States
TargetNo. 1
Index (1,70)
Lifetime (s) 𝟐
No. 2 (1,100) [400; − 𝐦/𝐬,
10; 0.4;
(𝐦, 𝐦/𝐬
−600; 5;0.01], 𝐦, 𝐦/𝐬, 𝐦/
No. 3 (1,70) [−800; 20; −0.16; −200; −5;0.3]
No.No.14 (1,70)
(20,100) [400; −7; −0.37; [0; −0;0.03; 0; −37;0,75]
100; −42;1]
No. 5 (20,100) [400; −1.4; −0.44; −600; 10; 0.05]
No.No.26 (1,100)
(20,100) [0; 5.5; [400; −10;0.4;
0.29; 0; 22; −0.55] −600; 5;0.01
No.No.37 (1,70)
(40,100) [−800; 32;[−800;20;
−1.5; −200; −0.16;
11; −0.75]−200; −5;0
No. 8 (40,100) [−200; 15; 0.3; 800; −10; 0]
No.No.49 (20,100)
(60,100) [−800; −3; [400; −7; 15;
1.5; −200; −0.37;
−0.75]−100; −42;
No.
No. 5 10 (60,100)
(20,100) [−200;
[400;−3; −1.4;
0; 800; − 15; 0] −600; 10; 0
−0.44;
No. 11 (80,100) [0; −20; −1.6; 0; −45; 0.2]
No.
No. 612 (20,100)
(80,100) [−200; 15;[0;0;5.5; 0.29;
800; − 5; 0] 0; 22; −0.55]
No. 7 (40,100) [−800; 32; −1.5; −200; 11; −0
As shown in Figures 12 and 13, the sector-matching framework proposed in this paper
No. 8 (40,100) [−200; 15; 0.3; 800; −10; 0
can achieve radar MTT when considering radar measurement time diversity. Applying
No.to9the LGM-PHD filter
the method (60,100)
can achieve track maintenance[−800;
even−3; 1.5;presence
in the −200; 15; −0.
of strong
No.clutter
10 interference, dense targets in the surveillance area[−200;
(60,100) and long−3;radar scan −15; 0]
0; 800;
periods. Similar to example 1, in Figure 12, a low detection probability of p D,k = 0.8
No.trajectory
causes the 11 to be broken,(80,100)
and in Figure 12b,d, the reason why[0; −20; −1.6; 0;can
the trajectory −45; 0.2]
No. 12
be identified as the same target’s (80,100) [−200;
trajectory instead of a new-born target 15; 0; 800;
trajectory after −5; 0]
the trajectory is broken is that in the simulation made in this paper, it is assumed that the
new-born targets can only appear at the known new born points, i.e., it is assumed that
The radar
no new-born divides
target the detection
will appear airspace
at the trajectory breaking into 500Inparts.
points. Figure A13,detailed
due to the compa
performance of the individual
extension of accumulation filters at
time, the detection 𝑇 = 1 (p
probability s,D,k𝑇 ==0.96)
2 ms, 𝜆 =probability
and the 104, 𝑝 , = 0
2 s, 𝑇 = 4 ms, 𝜆 = 52, 𝑝 , = 0.96, and the effect of different sampling times
of track breakage is significantly reduced.
In Figure 14, the computation time (CT) of the MATLAB simulation for 100 sample
performance
runs is given and arethe given
average in Figures
computation 12–16.
is shown in Table 4. The results in Figure 14
are As shown
the same as inin Figures
Figure 12 and
8, which show 13,
that the sector-matching
the real-time performances framework
of the proposed propose
method for both LGM-PHD filter and the basic GM-PHD
per can achieve radar MTT when considering radar measurement time diver filter are similar to that of the
basic PHD filter and the LGM-PHD filter, and much better than that of the joint GLMB
ingfilter.
theInmethod
addition,to the LGM-PHD
extending filter time
the accumulation can can
achieve
reduce track maintenance
the calculation time. even in t
of strong clutter
In Figures 15 interference,
and 16, the average denseOSPA targets
distanceinofthe
the surveillance
MATLAB simulation area forand long
100 sample runs is given and the average OSPA distances
periods. Similar to example 1, in Figure 12, a low detection probability o are given in Table 4. Compared
with the basic GM-PHD filter, the average tracking performance of the SM-LGM-PHD
causes theLGM-PHD
filter, the trajectory tothebeSM-GM-PHD
filter, broken, and filterin
andFigure 12b,d, filter
the joint-GLMB the reason why the tr
in this scenario
is improved by 41.57%, 16.40%, 28.78% and 47.67%, respectively, at Ts = 1 s (Ti = 2 ms,
λ = 104, p D,k = 0.8), and 57.32%, 29.11%, 34.66% and 37.69%, respectively, at Ts = 2 s
(Ti = 4 ms, λ = 52, p D,k = 0.96). Compared with the Ts = 1 s scenario, the performance
Appl. Sci. 2023, 13, 2834 18 of 22

of the SM-LGM-PHD filter and the SM-GM-PHD filter is highly improved in the Ts = 2 s
scenario. The LGM-PHD filter, the basic GM-PHD filter and the joint-GLMB filter are
not time-matched, which is why the performance of these three filters is not particularly
improved, and the performance of the joint-GLMB filter even decreases. In addition, it can
be seen in Figures 15b and 16b that the values of OSPA location of the LGM-PHD18filter,
Appl. Sci. 2023, 13, x FOR PEER REVIEW of 23
the basic GM-PHD filter and the joint-GLMB filter all increase with the accumulation of
time. After 80s, with the appearance of two targets, No.11 and No.12, at a higher speed,
the targets’ position estimation errors of the LGM-PHD filter, the basic GM-PHD filter
be identified
and as thefilter
the joint-GLMB sameincrease
target’sattrajectory instead
a faster rate, of the
while a new-born
SM-LGM-PHDtarget trajectory
filter and after
the
the trajectory is broken is that in the simulation made in this paper, it is
SM-GM-PHD filter are not significantly affected. This is mainly caused by the additional assumed that the
new-born targets
measurement error can onlybyappear
caused samplingat the known
time newmentioned
diversity born points, i.e., it is assumed that
earlier.
no new-born
Comparedtarget
with will appear
the CV at the
model trajectory1,breaking
in example points. model
the CA motion In Figurehas13, due to the
a quadratic
extension
term of so
in time, accumulation
the increases time, theestimation
in the detection probability
error due to(𝑝the
, = 0.96)
time and the
accumulation probability
of the
of track breakage is significantly reduced.
basic PHD filter, the LGM-PHD filter and the GLMB filter are more significant.
90° 90°
105° 75° 105° 75°
120° 60° 120° 60°

135° 45° 135° 45°

150° 30° 150° 30°

165° 15° 165° 15°


Estimates
Estimates
True tracks
True tracks
Measurements
Measurements

180° 0° 180° 0°
0m 400m 800m 1200m 1600m 2000m 0m 400m 800m 1200m 1600m 2000m

(a) (b)

(c) (d)
Figure12.
Figure 12.The
Theestimations
estimations ofof
thethe multi-target
multi-target trajectories
trajectories andand
thethe actual
actual measurements
measurements in example
in example 2
2 (CA motion model) for a sample run. (a) The actual measurements, actual trajectories and esti-
(CA motion model) for a sample run. (a) The actual measurements, actual trajectories and estimated
mated trajectories obtained after filtering by the SM-GM-PHD filter for 𝑇 = 1 s (𝑇 = 2 ms, 𝜆 =
trajectories obtained after filtering by the SM-GM-PHD filter for Ts = 1 s (Ti = 2 ms, λ = 104,
104, 𝑝 , = 0.8). (b) The actual measurements, actual trajectories and estimated trajectories ob-
p D,k = 0.8).
tained after(b)filtering
The actual measurements,
by the SM-LGM-PHD actual trajectories
filter for 𝑇 = and 1 s, estimated trajectories
and the different obtainedare
trajectories after
de-
filtering
noted byby different
the SM-LGM-PHD
color dots.filter for T
(c) The s = 1 measurements,
actual s, and the different trajectories
actual are denoted
trajectories by different
and estimated trajec-
color dots.
tories (c) Theafter
obtained actual measurements,
filtering actual trajectories
by the SM-GM-PHD filter inand estimated
x and trajectories
y coordinates = 1 s. (d)
for 𝑇obtained after
The
actual measurements,
filtering by the SM-GM-PHD actualfilter
trajectories
in x and and estimatedfor
y coordinates trajectories obtained
Ts = 1 s. (d) aftermeasurements,
The actual filtering by the
SM-LGM-PHD
actual trajectoriesfilter
and in x and y trajectories
estimated for 𝑇 = after
coordinatesobtained 1 s. filtering by the SM-LGM-PHD filter in x
and y coordinates for Ts = 1 s.
Appl.Sci.
Appl. Sci.2023,
2023,13,
13,2834
x FOR PEER REVIEW 19 ofof2223
19

90° 90°
105° 75° 105° 75°

120° 60° 120° 60°

135° 45° 135° 45°

150° 30° 150° 30°

165° 15° 165° 15°


Estimates Estimates
True tracks True tracks
Measurements Measurements
180° 0° 180° 0°
0m 400m 800m 1200m 1600m 2000m 0m 400m 800m 1200m 1600m 2000m

(a) (b)

(c) (d)
Figure13.
Figure 13.The
Theestimations
estimationsofofthe
themulti-target
multi-target trajectories
trajectories and
and thethe actual
actual measurements
measurements in example
in example 2
2 (CA motion model) for a sample run. (a) The actual measurements, actual trajectories and esti-
(CA motion model) for a sample run. (a) The actual measurements, actual trajectories and estimated
mated trajectories obtained after filtering by the SM-GM-PHD filter for 𝑇 = 2 s (𝑇 = 4 ms, 𝜆 = 52,
trajectories obtained after filtering by the SM-GM-PHD filter for Ts = 2 s (Ti = 4 ms, λ = 52,
𝑝 , = 0.96). (b) The actual measurements, actual trajectories and estimated trajectories obtained
D,k =filtering
pafter 0.96). (b)byThe
the actual measurements,
SM-LGM-PHD actual
filter for 𝑇 =trajectories
2 s, and theand estimated
different trajectories
trajectories obtained
are denoted by
after filtering by the SM-LGM-PHD filter
different color dots. (c) The actual measurements, for T s = 2 s, and the different trajectories are denoted
actual trajectories and estimated trajectories ob-
by different
tained aftercolor dots.by(c)the
filtering The actual measurements,
SM-GM-PHD filter in x andactual for 𝑇
trajectories and
y coordinates = 2 s. (d)
estimated trajectories
The actual
Appl. Sci. 2023, 13, x FOR PEER REVIEW
obtained after filtering
measurements, by the SM-GM-PHD
actual trajectories and estimated filter in x andobtained
trajectories y coordinates for Ts =
after filtering 20 of
2 s. SM-LGM-
by the (d) 23
The
PHD measurements,
actual filter in x and y coordinates for 𝑇 =
actual trajectories and2 s.estimated trajectories obtained after filtering by the
SM-LGM-PHD filter in x and y coordinates for Ts = 2 s.
In Figure 14, the computation time (CT) of the MATLAB simulation for 100 sample
runs is given and the average computation is shown in Table 4. The results in Figure 14
are the same as in Figure 8, which show that the real-time performances of the proposed
method for both LGM-PHD filter and the basic GM-PHD filter are similar to that of the
Computing time(s)

basic PHD filter and the LGM-PHD filter, and much better than that of the joint GLMB
filter. In addition, extending the accumulation time can reduce the calculation time.

(a) (b)
Figure 14.
Figure 14. Computation
Computation time
time of
of different
different filters
filters for
for 100
100 sample
sample runs at 𝑇
runs at Ts =
= 11 ss (T
(𝑇 = 2 ms, 𝜆 = 104,
i = 2 ms, λ = 104,
𝑝 , = 0.8) (a) and 𝑇 = 2𝑠 (𝑇 = 4 ms, 𝜆 = 52, 𝑝 , = 0.96) (b) in example 2 (CA motion model).
p D,k = 0.8) (a) and Ts = 2s (Ti = 4 ms, λ = 52, p D,k = 0.96) (b) in example 2 (CA motion model).

In Figures 15 and 16, the average OSPA distance of the MATLAB simulation for 100
sample runs is given and the average OSPA distances are given in Table 4. Compared
with the basic GM-PHD filter, the average tracking performance of the SM-LGM-PHD
filter, the LGM-PHD filter, the SM-GM-PHD filter and the joint-GLMB filter in this sce-
nario is improved by 41.57%, 16.40%, 28.78% and 47.67%, respectively, at 𝑇 = 1 s (𝑇 =
2 ms, 𝜆 = 104, 𝑝 , = 0.8), and 57.32%, 29.11%, 34.66% and 37.69%, respectively, at 𝑇 =
2 s (𝑇 = 4 ms, 𝜆 = 52, 𝑝 , = 0.96). Compared with the 𝑇 = 1 s scenario, the perfor-
Appl. Sci. 2023, 13, x FOR PEER REVIEW 21 of 23

Appl. Sci.Sci.
Appl. 2023, 13,13,
2023, x FOR
2834PEER REVIEW 21 of
20 23
of 22

OSPA Dist
OSPA Dist
Loc Loc
OSPAOSPA
CardCard
OSPA
OSPA

Figure 15. The average OSPA of different filters in example 2 (CA motion model) for 100 sample
The average OSPA
Figure
Figure 15.15.
runs when The𝑇 =average
1 s (𝑇OSPA of of
= 2 ms,
different filters
different
𝜆 = 104,filters
in
𝑝 , =in0.8).
example
example 2 (CA
2 (CA
(a) Average
motion model)
motion
OSPAmodel)
distancesforfor 100
100 sample
sample
of different fil-
runs
runs when
when 𝑇 =T = 1 s (T
s 1 s (𝑇 = = 2 ms, = 104,
i 2 ms, 𝜆 = 104, 𝑝 D,k
λ p = 0.8).
= OSPA (a) Average
0.8). (a)cardinalitiesOSPA
Average OSPA distances
distances of
of differentfilters.
different fil-
ters. (b) OSPA locations of different filters., (c) of different filters.
(b)(b)
ters. OSPA locations
OSPA of different
locations filters.
of different (c) (c)
filters. OSPA cardinalities
OSPA of different
cardinalities filters.
of different filters.
OSPA
OSPA Dist
Dist
Loc
OSPA Loc
OSPA Card
OSPACard
OSPA

Figure
Figure16.16.
The
Theaverage
average OSPA
OSPA of of
different filters
different in in
filters example
example 2 (CA
2 (CA motion
motionmodel)
model)forfor
100 sample
100 sample
Figure 16. The average
when𝑇 T= 2 s2 s(𝑇(T=OSPA
4 4ms, of different
ms,𝜆λ==52,
filters
52,𝑝p D,k
in example 2 (CA motion model) for 100filters.
sample
runs, when
runs, s = i = , ==0.96).
0.96).(a)
(a)Average
Average OSPA
OSPA distances ofof different
different fil-
runs,(b)
ters. when
OSPA 𝑇 locations
= 2 s (𝑇of=different
4 ms, 𝜆 filters.
= 52, (c)
𝑝 ,OSPA= 0.96). (a) Average
cardinalities of OSPA filters.
different distances of different fil-
(b) OSPA locations of different filters. (c) OSPA cardinalities of different filters.
ters. (b) OSPA locations of different filters. (c) OSPA cardinalities of different filters.
Appl. Sci. 2023, 13, 2834 21 of 22

Table 4. List of the average CT and average OSPA of each filter in example 2 (CA model).

SM-LGM- LGM- SM-GM- Basic GM- Joint-


PHD Filter PHD Filter PHD Filter PHD Filter GLMB Filter
CT ( Ts = 1 s ) 0.0248 s 0.0223 s 0.0210 s 0.0174 s 0.0956 s
CT ( Ts = 2 s ) 0.0151 s 0.0133 s 0.0178 s 0.0160 s 0.1038 s
OSPA ( Ts = 1 s ) 23.98 m 34.31 m 29.23 m 41.04 m 21.47 m
OSPA ( Ts = 2 s ) 16.76 m 27.84 m 25.66 m 39.27 m 24.47 m

5. Conclusions
A computationally efficient sector-matching PHD filter is proposed in this paper.
With the radar sampling time diversity considered, the proposed filter is well adapted to
the MTT problem during radar dynamic scanning. The proposed sector-matching struc-
ture quantifies the relationship between beam dwell time, false alarm rate and detection
probability, and divides the scanning area into small sectors to obtain actual multi-target
measurement times. The proposed filter rederives the prediction and updating steps based
on the actual sampling time, and a state correction step is added. The performance is
evaluated by simulations using the Gaussian linear model, and the results demonstrate
that the proposed filter provides enhanced filtering performance in multi-target tracking.
There are still some challenges with the proposed filter, including radars operating in TAS
mode and target migration through resolution cells. We will address these issues in future
research. In addition, as the performance of the TPHD [33] and TPMB [23] filters is much
better than the basic PHD filter, future work will investigate the combination of “sector
matching” structures with trajectory filters.

Author Contributions: Conceptualization, J.Y.; formal analysis, J.Y. and J.T.; investigation, D.J. and
Y.G.; resources, D.J. and M.L.; data curation, J.Y. and Y.H.; writing—original draft preparation, J.Y.;
writing—review and editing, J.Y., D.J., Y.G., X.L. and M.L. All authors have read and agreed to the
published version of the manuscript.
Funding: This research was funded by the National Natural Science Foundation of China under
grant no. 61971179 and the Fundamental Research Funds for the Central Universities under grant
no. B200202165.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Mahler, R.P.S. Statistical Multisource-Multitarget Information Fusion; Artech House: Norwood, MA, USA, 2007.
2. Mahler, R.P.S. Advances in Statistical Multisource-Multitarget Information Fusion; Artech House: Norwood, MA, USA, 2014.
3. Pace, P.E. Detecting and Classifying Low Probability of Intercept Radar; Artech House: Norwood, MA, USA, 2008.
4. Fu, W.; Jiang, D.; Su, Y.; Qian, R.; Gao, Y. Implementation of wideband digital transmitting beamformer based on LFM waveforms.
IET Signal Process. 2017, 11, 205–212. [CrossRef]
5. Jiang, Q.; Wang, R.; Zhou, C.; Zhang, T.; Hu, C. Modified Bayesian Group Target Track Initiation Algorithm Based on Algebraic
Graph Theory. J. Electron. Inf. Technol. 2021, 43, 531–538.
6. Granstrom, K.; Lundquist, C.; Gustafsson, F.; Orguner, U. Random Set Methods Estimation of Multiple Extended Objects. IEEE
Robot. Autom. Mag. 2014, 21, 73–82. [CrossRef]
7. Fortmann, T.; Bar-Shalom, Y.; Scheffe, M. Sonar tracking of multiple targets using joint probabilistic data association. IEEE J.
Ocean. Eng. 1983, 8, 173–184. [CrossRef]
8. Chong, C.Y. Tracking and data fusion: A handbook of algorithms (bar-shalom, y. et al.; 2011)[bookshelf]. IEEE Control. Syst. Mag.
2012, 32, 114–116. [CrossRef]
9. Blackman, S. Multiple hypothesis tracking for multiple target tracking. IEEE Aerosp. Electron. Syst. Mag. 2004, 19, 5–18. [CrossRef]
10. Wang, S.; Bao, Q.; Chen, Z. Refined PHD Filter for Multi-Target Tracking under Low Detection Probability. Sensors 2019, 19, 2842.
[CrossRef]
Appl. Sci. 2023, 13, 2834 22 of 22

11. Chai, L.; Yi, W.; Jiang, X.; Kong, L. A Distributed PHD Filter for On-line Joint Sensor Registration and Multi-target Tracking. In
Proceedings of the International Conference on Information Fusion, Ottawa, ON, Canada, 2–5 July 2019; pp. 1–6.
12. Li, T.; Mallick, M.; Pan, Q. A Parallel Filtering-Communication-Based Cardinality Consensus Approach for Real-Time Distributed
PHD Filtering. IEEE Sens. J. 2020, 20, 13824–13832. [CrossRef]
13. Jiang, D.; Liu, M.; Gao, Y.; Gao, Y.; Fu, W.; Han, Y. Time-Matching Random Finite Set-Based Filter for Radar Multi-Target Tracking.
Sensors 2018, 18, 4416. [CrossRef]
14. Clark, D.; Vo, B.T.; Bell, J. GM-PHD filter multi-target tracking in sonar images. In Proceedings of the Conference on Signal
Processing, Sensor Fusion, and Target Recognition XV, Kissimmee, FL, USA, 17–19 April 2006. [CrossRef]
15. Clark, D.; Ruiz, I.; Petillot, Y.; Bell, J. Particle PHD filter multiple target tracking in sonar image. IEEE Trans. Aerosp. Electron. Syst.
2007, 43, 409–416. [CrossRef]
16. Moratuwage, D.; Vo, B.N.; Wang, D.W. Collaborative Multi-Vehicle SLAM with Moving Object Tracking. In Proceedings of the
IEEE International Conference on Robotics and Automation (ICRA), Karlsruhe, Germany, 6–10 May 2013.
17. Hoseinnezhad, R.; Vo, B.N.; Vo, B.T.; Suter, D. Visual tracking of numerous targets via multi-Bernoulli filtering of image data.
Pattern Recognit. 2012, 45, 3625–3635. [CrossRef]
18. Pollard, E.; Plyer, A.; Pannetier, B.; Champagnat, F.; Le Besnerais, G. Gm-phd filters for multi-object tracking in uncalibrated
aerial videos. In Proceedings of the 12th International Conference on Information Fusion, Seattle, WA, USA, 6–9 July 2009; IEEE:
New York, NY, USA, 2009; pp. 1171–1178.
19. Vo, B.N.; Ma, W.K. The Gaussian Mixture Probability Hypothesis Density Filter. IEEE Trans. Signal Process. 2006, 54, 4091–4104.
[CrossRef]
20. Vo, B.T.; Vo, B.N. Labeled Random Finite Sets and Multi-Object Conjugate Priors. IEEE Trans. Signal Process. 2013, 61, 3460–3475.
[CrossRef]
21. Reuter, S.; Vo, B.T.; Vo, B.N.; Dietmayer, K. The Labeled Multi-Bernoulli Filter. IEEE Trans. Signal Process. 2014, 62, 3246–3260.
22. Vo, B.N.; Vo, B.T.; Hoang, H. An Efficient Implementation of the Generalized Labeled Multi-Bernoulli Filter. IEEE Trans. Signal
Process. 2017, 65, 1975–1987. [CrossRef]
23. García-Fernández, Á.F.; Svensson, L.; Williams, J.L.; Xia, Y.; Granström, K. Trajectory Poisson Multi-Bernoulli Filters. IEEE Trans.
Signal Process. 2020, 68, 4933–4945. [CrossRef]
24. Meyer, F.; Kropfreiter, T.; Williams, J.L.; Lau, R.; Hlawatsch, F.; Braca, P.; Win, M.Z. Message passing algorithms for scalable
multitarget tracking. Proc. IEEE 2018, 106, 221–259. [CrossRef]
25. Gan, R.Z.; Li, Q.; Godsill, S. A Variational Bayes Association-based Multi-object Tracker under the Non-homogeneous Poisson
Measurement Process. In Proceedings of the 2022 25th International Conference on Information Fusion (FUSION), Linköping,
Sweden, 4–7 July 2022; IEEE: New York, NY, USA, 2022; pp. 1–8.
26. Khan, Z.; Balch, T.; Dellaert, F. Mcmc Data Association and Sparse Factorization Updating for Real Time Multitarget Tracking
with Merged and Multiple Measurements. IEEE Trans. Pattern Anal. Mach. Intell. 2006, 28, 1960–1972. [CrossRef]
27. Davey, S.J.; Fabrizio., G.A.; Rutten, M.G. Detection and tracking of multipath targets in over-the-horizon radar. IEEE Trans. Aerosp.
Electron. Syst. 2019, 55, 2277–2295. [CrossRef]
28. Streit, R.; Degen, C.; Koch, W. The pointillist family of multitarget tracking filters. arXiv 2015, arXiv:1505.08000.
29. Lin, L.; Bar-shalom, Y.; Kirubarajan, T. Tracking labeling and PHD filter for multitarget tracking. IEEE Trans. Aerosp. Electron. Syst.
2006, 42, 778–795. [CrossRef]
30. Yang, J.; Ji, H. A novel track maintenance algorithm for PHD/CPHD fifilter. Signal Process. 2012, 92, 2371–2380. [CrossRef]
31. Li, Y.; Xiao, H.; Wu, H.; Fu, Q.; Hu, R. Modified Labeled Particle Probability Hypothesis Density Filter for Joint Multi-target
Tracking and Classification. In Proceedings of the International Conference on Signal Processing and Communication Systems,
Cairns, QLD, Australia, 14–16 December 2015; pp. 1–5.
32. Li, T.; Sun, S.; Corchado, J.M.; Siyan, M.F. A particle dyeing approach for track continuity for the SMC-PHD filter. In Proceedings
of the International Conference on Information Fusion, Slamanca, Spain, 7–10 July 2014; pp. 1–8.
33. García-Fernández, Á.F.; Svensson, L. Trajectory probability hypothesis density filter. In Proceedings of the 2018 21st International
Conference on Information Fusion (FUSION), Cambridge, UK, 10–13 July 2018; pp. 1430–1437.
34. García-Fernández, Á.F.; Svensson, L. Trajectory PHD and CPHD Filters. IEEE Trans. Signal Process. 2019, 67, 5702–5714. [CrossRef]
35. Gao, Y.; Jiang, D.; Zhang, C.; Guo, S. A Labeled GM-PHD Filter for Explicitly Tracking Multiple Targets. Sensors 2021, 21, 3932.
[CrossRef]
36. Zhang, Y.; Liu, M.; Zhao, K.; Gao, Y.; Zhang, T. Time-matching Recursive Extended Target Probability Hypothesis Density Filter
for High Resolution Phased Array Radar. In Proceedings of the 2021 33rd Chinese Control and Decision Conference (CCDC),
Kunming, China, 22–24 May 2021; IEEE: New York, NY, USA, 2021; pp. 7218–7223.
37. Jiang, D.F.; Liu, M.; Gao, Y.Y.; Gao, Y. Time-matching extended target probability hypothesis density filter for multi-target tracking
of high resolution radar. Signal Process. 2019, 157, 151–160. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.

You might also like