2318 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 9, NO. 4, NOVEMBER 2013
An Efficient Data-Driven Particle PHD Filter for
Multitarget Tracking
Yunmei Zheng, Student Member, IEEE, Zhiguo Shi, Member, IEEE, Rongxing Lu, Member, IEEE,
Shaohua Hong, Member, IEEE,and XueminShen, Fellow, IEEE
Abstract—In this paper, we propose an efficient data-d
riven par-
ticle probability hypothesis density (PHD) filter for real-time mul-
titarget tracking of nonlinear/non-Gaussian system in dense clutter
environment. In specific, the input measureme
nts are first classi-
fied into two sets, namely survival measurements and spontaneous
birth measurements, after e liminating clutters by using existing
historic state d ata of targets. Since most c
lutters do not participate
in the complex weight computation of particle PHD filter, better
real-time performance can b e achieved. The tracking p erformance
is also improved because the survival me
asurements are used for
survival targets and the spontaneous birth measurements are used
for spontaneous birth targets, resulting in less interference from
each other and from clutters. Exte
nsive simulations validate the
improvement of both the real-time performance and tracking per-
formance of the proposed data-driven particle PHD filter in com-
parison with the traditional pa
rticle PHD filter.
Index Terms—Data-driven mechanism, particle probability
hypothesis density (PHD) filter, real-time performance, tracking
performance.
I. INTRODUCTION
M
ULTIPLE target tracking (MTT) is a very important
technology for many industrial applicat ions, such as
automated surveillance [1], wireless sensor networks [2]–[4],
mobile robots [5], traffic monitoring [6], etc. Recently, the
so-called probability hypo thesis density (PHD) filter and cardi-
nalized PHD (CPHD) filter which avoid exp licit associations
between measurements and targets have been widely studied
for MTT problems. The idea of PHD/CPHD filter is to repr e sent
the targets and measurements as random finite sets (RFSs) and
use fin ite set stati stics (FISST) to solve MT T problems under a
Bayesian framework.
Manuscript received March 01, 2012; revised June 09, 2012, September 20,
2012; accepted October 30, 2012. Date of publication November 21, 2012; date
of current version October 14, 2013. This work was supported in part by the
National Science Foundation of China under Grant 61171149, the Zhejiang
Province Commonweal Technique Research Project 2010C31069, the Research
Foundation of State Key Laboratory of Industrial Control Technology under
Grant ICT1119, and ORF-RE, Ontario, Canada. Paper no. TII-12-0128.
Y. Zheng and Z. Sh i a re with the Departme nt of Information S cience
and Electronic Engineering, Zhejiang University, Hangzhou, 310027, China
(e-mail: zhengyunmei@zju.edu.cn; shizg@zju.edu.cn).
R. Lu a nd X. Shen a re with the D epartm ent of Electrical and Com-
puter Engineering, University of Waterloo, Waterloo, Canada (e-mail:
rxlu@bbcr.uwaterloo.ca; xshen@bbcr.uwaterloo.ca).
S. Hong is with the Dep artment of Communication Engineerin g , Xiamen Uni-
versity, Fujian, 361005, China (e-mail: hongsh@xmu.edu.cn).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Dig
ital Object Identifier 10.1109/TII.2012.2228875
For the PHD filter [7], it propagates the intensity of th
eRFS
of states in time, the advantage of which is that it ope
rates only
on the single-target state space and complete
ly avoids any data
association computation. For the CPHD filter
[8], [9], it propa-
gates the intensity of the RFS and the e ntire
probability distri-
bution of the target number in time, whi ch re
laxes the Poi s son
distribution assumption on the number of t
argets in the PHD
filter at the cost of much higher computati
onal complexity than
that of the PHD filter [10]. From imple
mentation p erspective,
a full sequential M onte Carlo (SMC)
implementation of PHD
filter, also called particle PHD fil
ter, was proposed in [11], and
closed-form solutions to the PHD
/CPHD recursions were de-
rived for linear Gaussian multi
target models in [12] and [13],
respectively.
In addition to the PHD/CPHD filte
r, Mahler has proposed the
multitarget multi-Bernoul
li (MeMBer) recursion as a tractable
approximation to the Bayes
multitarget r ecurs ion under low-
clutter-density scenario
s [14]. Unlike the PHD/CPHD recur-
sions, the MeMBer recursi
on propagates (approximately) the
multitarget posterior d
ensity, and it allows reliable and inex-
pensive extraction of st
ate estimates without clustering in the
PHD/CPHD filter.
The demands of “real-t
ime” MTT have been increasing
[15]–[20]. Since the
CPHD filter propagates b oth the intensity
of th e RFS and the post
erior cardinality distribution [13], its
real-time characte
ristic is intrin sicall y not as good as the PHD
filter. Althou gh GM
-PHD/CPHD has a closed-form solution
which makes it easy
for r eal-tim e implementation, its applica-
tion scenario is
constrained to linear Gaussian system . W hen
comparing the P
HD filter with the MeMBer filter, according
to the modeling
assumptions on the PHD filter [11] and the
MeMBer filter [
21], the PHD filter is more suitable for denser
clutter envi
ronments. Thus, we consider particle PHD filter as
a good candid
ate for nonlinear/non-Gaussian MTT problems
in dense c l
utter scenarios with high real-time requirements.
However, s
ince the particle PHD fi
lter is a kind of SMC ap-
proach, i
ts computational complexity is very high. Therefore,
we are int
erested in improving the real-time performance of the
particl
ePHDfilter.
Similar t
o the particle filter [22]–[24], the parti cle PHD
filter ma
inly consists of three steps, namely, Generation of
Partic
les (Prediction), Weight Computation (Update), and
Resamp
ling (Resample). Specifically, Weight Computation
consi
sts of a mass of complicated m athematical computations
and fo
rms a major bottleneck of the traditional particle PHD
proc
essing. Since the measurements act as the only i nput in the
par
ticle PHD filter, it is po ssible to use data-driven approach
1551-3203 © 2012 IEEE