Journal of Systems Engineering and Electronics
Vol. 26, No. 4, August 2015, pp.1 – 9
Immune adaptive Gaussian mixture particle filter for
state estimation
Wenlong Huang
*
, Xiaodan Wang, Yi Wang, and Guohong Li
Air and Missile Defense College, Air Force Engineering University, Xi’an 710051, China
Abstract:
The particle filter (PF) is a flexible and powerful sequen-
tial Monte Carlo (SMC) technique capable of modeling nonlinear,
non-Gaussian, and nonstationary dynamical systems. However,
the generic PF suffers from particle degeneracy and sample im-
poverishment, which greatly affects its performance for nonlinear,
non-Gaussian tracking problems. To deal with those issues, an
improved PF is proposed. T he algorithm consists of a PF that
uses an immune adaptive Gaussian mixture model based immune
clone algorithm to re-approximate the posterior density. At the
same time, three immune antibody operators are embed in the
new filter. Instead of using a resample strategy, the newest obser-
vation and conditional likelihood are integrated into those immune
antibody operators to update the particles, which can further im-
prove the diversity of particles, and drive particles toward their
close local maximum of the posterior probability. The improved PF
algorithm can produce a closed-form expression for the posterior
state distribution. Simulation results show the proposed algorithm
can maintain the effectiveness and diversity of particles and avoid
sample impoverishment, and its performance is superior to several
PFs and Kalman filters.
Keyw ords: artificial immune, particle filter, Gaussian mixture
model.
DOI: 10.1109/JSEE.2015.00001
1. Introduction
The par ticle filter (PF) is an attractive estimation proce-
dure for non-linear dynamical systems [1]. It is proven
to be very successful for solving non-linear and non-
Gaussian state estimation problems, including target trac-
king, speech recognition, financial econometrics, and mo-
bile robot [2 – 6]. In the generic PF, the algorithms appro-
ximate the posterior densities of the hidden states with a
Manuscript received April 14, 2014.
*Corresponding author.
This work was supported by the National Natural Science Founda-
tion of China (61273275; 61402517), the Open Research Fund of State
Key Laboratory of Astronautic Dynamics (2012ADL-DW0202), the Na-
tural Science Foundation of Shaanxi Province of China (2013JQ8035),
and the project funded by China Postdoctoral Science Foundation
(2013M542331).
set of p articles, and the particles are sampled from the state
space and the posterior is updated by propagating the par-
ticles and updating their weights based upon the observa-
tions’ likelihoods. Generally, the performance of the PF is
superior to other non-linear filters [7]. In principle, an in-
finite number of particles can exactly represent any given
probability density function (PDF). However, the require-
ment for the number of particles grows exponentially with
dimension, which makes PF suffer due to ‘curse of dimen-
sionality’. This problem is exacerbated by the high dimen-
sionality of the state space [8 – 10].
Moreover, the sample degeneracy phenomenon will un-
fortunately unavoidablly occur in PF, when there ar e only
a few particles in the vicin ity of the true state. Only these
particles have significant weights, and the weights of most
of the other particles are near zero. The posterior density
cannot be approximated with small sample sets properly,
and significant computation is wasted on particles with low
weights. In order to reduce the degeneracy effect, several
related algorithms h ave been introduced that incorporate
density estimation methods into a particle filter [10 – 14].
In [15], Cham and Rehg employed a piecewise Gaussian
model to represent the measurement function and the pre-
diction density. In [1 6], Vermaak et al. directly modeled
the posterior density as a Gaussian mixture model (GMM).
A kernel-based Bayesian filter was introduced in [17]. An
analytical form for the measurement function p(z
k
|x
1:k
)
is estimated from the data using multistage sampling and
density interpolation m ethods. After multiplying the two
functions, a smooth approximation for the posterior den-
sity is found using a variable-width mean-shift technique.
While these methods exploit an analytical form to main-
tain multiple hypotheses or simplify the sampling problem,
PFs suffer from a high computational cost, especially when
large numbers of particles are required to model the poste-
rior distribution.
On the other hand, during the resampling step, parti-
cles having large weights are repeatedly selected and the