Performance Comparison of Several Nonlinear
Multi-Bernoulli Filters for Multi-Target Filtering
Meiqin Liu, Tongyang Jiang, Xie Wang, and Senlin Zhang
Dept. of Systems Science and Engineering
College of Electrical Engineering, Zhejiang University
Hangzhou 310027, P. R. China
Email: {liumeiqin,jiangtongyang,wangxiek,slzhang}@zju.edu.cn
Abstract—In this paper, the performance of four nonlinear
multi-Bernoulli (MB) filters for multi-target filtering is compared
in the presence of clutter and detection uncertainty. The filters
under consideration are the extended Kalman (EK) Gaussian
mixture (GM) MB filter, the unscented Kalman (UK) GM-
MB filter, the cubature Kalman (CK) GM-MB filter, and the
sequential Monte Carlo (SMC) MB filter. Monte Carlo (MC)
analyses are presented for these four filters under different clutter
density and different detection probability. Then these filters are
evaluated in terms of both the Optimal Sub-Pattern Assignment
(OSPA) distance and their respective computing time. Simulation
results show that the CK-GM-MB filter is an attractive nonlinear
MB filtering approach.
I. INTRODUCTION
The random finite set (RFS) based multi-target filtering
methods have attracted more attention in recent years. The
finite set statistics (FISST) provides a rigorous Bayesian
framework for multi-target filtering [1]. Since the optimal
multi-target Bayes filter based on the RFS is generally
intractable, some sub-optimal multi-target Bayes filters are
proposed, such as the probability hypothesis density (PHD)
filter based on the first order moment approximation [2],
the cardinalized probability hypothesis density (CPHD) filter
based on the RFS moment and cardinality approximations [3],
and the multi-target multi-Bernoulli (MeMBer) filter based
on the RFS density approximations [1]. Since the MeMBer
filter overestimates the number of targets, Vo [4] improved
the MeMBer filter and proposed the cardinality balanced
Multi-target Multi-Bernoulli (CBMeMBer) filter which has
an unbiased estimation in the number of targets. Both the
Gaussian Mixture (GM) and sequential Monte Carlo (SMC)
techniques have been implemented for the PHD, CPHD,
and CBMeMBer filters [4],[5],[6],[7],[8]. The main advantage
of the CBMeMBer filter is that each Bernoulli distribution
represents one track and it’s existence probability, so it allows
reliable and inexpensive extraction of the target states in
the SMC implementation. The SMC-CBMeMBer filter also
outperforms the SMC-PHD and SMC-CPHD filters [4]. For
the GM implementation, the GM-CBMeMBer filter shows the
similar performance to the PHD filter.
Various implementations and extensions of the RFS-
based multi-Bernoulli filter have been considered in
[9],[10],[11],[12],[13],[14],[15]. The multi-Bernoulli filter has
been applied to tracking from audio and video information
[16], tracking in networks [17], tracking from image infor-
mation [18],[19], and sensor management for multi-target
tracking [20]. Hybrid multi-Bernoulli and Possion multi-target
filters have been proposed in [21]. The CBMeMBer filter will
be treated as the multi-Bernoulli (MB) filter throughout this
paper.
The GM-MB filter has an analytical solution under linear
Gaussian models, and can be extended to nonlinear models
using the extended Kalman (EK) or unscented Kalman (UK)
approximation. The SMC-MB filter can accommodate non-
linear and non-Gaussian models. The EK-GM-MB, UK-GM-
MB, and SMC-MB filters have been implemented for non-
linear multi-target filtering [8]. However, the comprehensive
comparison of these three mu lti-target filters for nonlinear
models is not performed in [8]. Recently the cubature Kalman
(CK) filter [22] which employs a third-degree spherical-radical
cubature rule was proposed to handle the nonlinear models.
Compared with the UK filter, the CK filter is more accurate
and more principled in mathematical terms [22]. Therefore, in
this paper, we study the joint usage of the CK approximation
together with the GM-MB filter to hand the nonlinear multi-
target filtering problem, and also compare its performance with
the EK-GM-MB, UK-GM-MB, and SMC-MB filters under
different scenarios. Then the Optimal Sub-Pattern Assignment
(OSPA) distance [23] and the computing time are used to
evaluate the performance of these four nonlinear MB filters.
The rest of this paper is organized as follows. Section II
provides the background on the MB-RFS and MB recursion.
The CK-GM-MB filter is studied in section III. In section
IV, the EK-GM-MB, UK-GM-MB, CK-GM-MB, and SMC-
MB filters are evaluated under different simulation conditions.
Finally, the conclusion is draw in Section V.
II. B
ACKGROUND
A. Multi-Bernoulli RFS
A Bernoulli RFS [4] has p robability of 1 − r of being
empty, and prob a bility r of being a singleton whose element
is distributed according to a probability p. The probability
density of a Bernoulli RFS X is given by
π(X)=
1 − rX= ∅,
r · p(x) X = {x}.
(1)