Gaussian Mixture Multiple-Model Multi-Bernoulli
Filters for Nonlinear Models Via Unscented
Transforms
Tongyang Jiang
∗†
,MeiqinLiu
∗†
,XieWang
†
, and Senlin Zhang
†
∗
State Key Laboratory of Industrial Control Technology, Zhejiang University, Hangzhou 310027, P.R. China
†
College of Electrical Engineering, Zhejiang University, Hangzhou 310027, P.R. China
Email: {jiangtongyang,liumeiqin,wangxiek,slzhang}@zju.edu.cn
Abstract—The multiple-model multi-Bernoulli (MM-MB) filter
is a new attractive approach for estimating multiple maneuvering
targets in the presence of clutter, missed detection and data
association uncertainty. In this paper, we extend the Gaussian
Mixture (GM) MM-MB filter to nonlinear models by using
unscented transform techniques. Moreover, in order to improve
the robustness and numerical stability of the unscented Kalman
(UK) GM-MM-MB filtering algorithm, we propose the square-
root UK (SUK) GM implementation of the MM-MB filter for
nonlinear models. A numerical example is presented to verify
the effectiveness of the UK-GM-MM-MB and SUK-GM-MM-
MB filtering approaches. Simulation results also show that the
SUK-GM-MM-MB filtering approach produces the same filtering
accuracy as the UK-GM-MM-MB filtering approach.
I. INTRODUCTION
Random finite set (RFS) based multi-target tracking ap-
proaches have attracted more attention in recent years. The
RFS-based approach treats the multi-target states and measure-
ments as RFSs, and jointly estimates the number of targets
and their states from the measurements. With RFS models,
Mahler has proposed the optimal multi-target Bayes filter that
propagates the posterior multi-target density recursively in
time [1],[2]. However, since the optimal multi-target Bayes
filter is generally intractable, some approximated approaches
have been proposed, such as the probability hypothesis density
(PHD) based on the first order moment approximation of
multi-target density [1], the cardinalized PHD (CPHD) filter
based on the moment and cardinality approximations [3],
and multi-target multi-Bernoulli (MeMBer) filter based on
density approximations [2]. Since the Mahler’s MeMBer filter
overestimates the number of target, Vo improved the MeMBer
filter and proposed a new version of the MeMBer filter called
the cardinality balanced (CB) MeMBer (CBMeMBer) filter
which has an unbiased estimation in the number of targets
[4]. The PHD, CPHD, and CBMeMBer filters have been
implemented by using Gaussian mixture (GM) and sequential
Monte Carlo (SMC) techniques [4],[5],[6],[7],[8]. In the SMC
implementation, the CBMeMBer filter has a reliable and
inexpensive extraction of target states, since it does not need
an extra clustering algorithm for extracting target states [5].
In the GM implementation, the CBMeMBer filter shows the
similar filtering performance to the PHD filter, and has a
lower computation complexity than the CPHD filter [4]. The
CBMeMber filter will be treated as the multi-Bernoulli (MB)
filter throughout this paper.
For maneuvering target tracking, the multiple-model (MM)
(or jump Markov system model) approach has proven to be an
effective method [9]. By integrating the MB filter with MM
approach, the MM-MB filter for maneuvering targets has been
proposed in [10],[11]. The GM implementation of MM-MB
filter for linear Gaussian models and the SMC implementation
of MM-MB filter for nonlinear models were also proposed in
[10],[11]. In target tracking, nonlinear models are commonly
used, such as radar and sonar measurements are nonlinear
[12]. Although the SMC-MM-MB filter can handle nonlinear
models, it still has some disadvantages. Firstly, similar to the
MM particle filter [13], the number of particles is proportional
to the model probability, if the model probability is very
low, only a small number of particles persists in the model.
This may cause filtering divergence. Secondly, although the
resampling step can reduce the degeneracy problem, it also
causes the loss of diversity among the particles, as the particles
after resampling step contain many repeated points. This
phenomenon will be severe if the process noise is small.
One solution to these problems is to increase the number of
particles. However, a large number of particles means a large
amount of calculation.
The GM-MM-MB filter has a close-form solution under
assumptions of linear Gaussian models. However, the GM-
MM-MB filter does not directly accommodate to nonlinear
models. In addition, at present there is no closed form solution
to GM-MM-MB filter for nonlinear models. Therefore, in this
paper we extend the GM-MM-MB filter to nonlinear models
by using unscented transforms [14]. Moreover, in order to im-
prove the robustness and numerical stability of the unscented
Kalman (UK) GM-MM-MB filter for nonlinear models, we
propose the square-root UK (SUK) GM implementation of the
MM-MB filter for nonlinear models. A numerical example is
also presented to compare the UK-GM-MM-MB and SUK-
GM-MM-MB filtering approach with the existing SMC-MM-
MB filtering approach.
The rest of this paper is organized as follows. Section II
provides the background on the MM-MB filter. The detailed
description of the UK-GM-MM-MB and SUK-GM-MM-MB
filtering approaches for nonlinear models are provided in
18th International Conference on Information Fusion
Washington, DC - July 6-9, 2015
978-0-9824-4386-6/15/$31.00 ©2015 IEEE 1262