Improved Sampling Efficiency in Particle Filter for Sys-
tems with Multi-Step Randomly Delayed Measurements
Xiaodan Chen, Junyi Zuo
*
, Hang Zou and Ti Zhang
School of Aeronautics
Northwestern Polytechnical University
Xi’an, Shaanxi Province, China
897135377@qq.com; junyizuo@nwpu.edu.cn; {2290249186, zhangti112358}@qq.com
Abstract - In particle filter, incorporating the current meas-
urement into the sampling process is an efficient strategy to im-
prove the sampling efficiency.
However, in the case of measure-
ment random delay, this strategy will become invalid since the
current received measurement is not necessary the one that the
sensor collects currently. To address this problem, we first calcu-
late the posterior distribution of the delay step of each sensor.
According to the calculation results and a predefined decision
logic, we can determine which measurements do not undergo
delay. Then the undelayed measurements are used to generate the
proposal distribution. Simulation results demonstrate the im-
proved performance of the proposed particle filter comparing
with the existing particle filters.
Index Terms - particle filter; proposal distribution; delayed
measurements.
I. INTRODUCTION
State estimation problem for nonlinear stochastic dynamic
systems with randomly delayed measurements has been gain-
ing more and more attention due to its wide existence in many
applications, such as target tracking, fault diagnosis and satel-
lite navigation. Many efforts have been devoted to this prob-
lem. In [1,2], an improved extended Kalman filter and an im-
proved unscented Kalman filter for nonlinear stochastic sys-
tems with one-step or two-steps randomly delayed measure-
ments have been proposed. A Gaussian filter and a Gaussian
smoother for nonlinear stochastic systems with one-step ran-
domly delayed measurements are proposed in [3] and [4],
respectively. These methods are only suitable for systems with
one-step or two-steps randomly delayed measurements. Re-
cently, a particle filter (PF) for nonlinear systems with multi-
ple step randomly delayed measurements is proposed in [5].
In this method, particles and their weights are updated by us-
ing the multiple step randomly delayed measurements in the
Bayesian estimation framework.
As we all know, particle filters have wide applications in
nonlinear filtering fields, which is attributed to its property of
no restriction to the linear and Gaussian assumptions. One of
the key steps of PF algorithm is to design an appropriate im-
portance density function (IDF), from which high-quality par-
ticles can be generated. In the standard PF, also called the
sampling importance resampling (SIR) filter [6], the transition
density is selected as the IDF. However, the transition density
only contains the model information, but ignores the infor-
*
The corresponding author is Pro. Zuo.
mation of the current measurements, which leads to low sam-
pling efficiency in some cases. Incorporating the current
measurement information into sampling process is an efficient
way to solve this problem. Based on this idea, many improved
particle filters have been proposed such as the auxiliary parti-
cle filter (APF) [7], the adaptive iterated particle filter (AIPF)
[8] and the unscented particle filter (UPF) [9]. These methods
tend to yield more precise results than SIR especially for the
case of low measurement noise level. However, in the pres-
ence of measurement random delay, directly incorporating the
current measurements information is problematic, since the
current received measurement may not be the one collected
by the sensor at current time. In fact, in this situation, using
the measurements to improve the sampling efficiency is a
challenging problem, which has not been investigated.
In this paper, we calculate the posterior probability of the
delay step of the current received measurements for each
sensor, and then decide whether the current received meas-
urement of each sensor undergoes delay or not through a pre-
defined decision logic. According to the decision results, we
select the undelayed measurements to generate the proposal
density. Thus, a new PF for nonlinear systems with multi-step
randomly delayed measurements (MSRDM) is derived. Sim-
ulation results show that the proposed method has higher es-
timation accuracy than many existing methods.
II. P
ROBLEM STATE MENT
Consider the following discrete-time nonlinear stochastic
system with
sensors as described by the following system
equation
()
-1 -1 -1ttt t
=+
fx v
(1)
and the MSRDM model
()
,, ,
,1,2,,
tk tk t tk
kK=+=zhxw
(2)
()
,,,
0
,1,2, 2
k
S
d
tk tk t dk
d
rk Kt
−
=
==≥
yz
(3)
where
t
is the discrete-time index,
t
x
is the state vector,
,1 , 2 ,
,,,
ttt tK
=zzz z
denotes the ideal measurement vector
received by the
sensors, and
,1 , 2 ,
,,,
ttt tK
=yyy y
is
the actual received measurement of the filter. Process noise
1t −
v
and the measurement noise
,tk
w
are zero-mean random
variables with known probability density functions (PDFs)
978-1-5386-8069-8/18//$31.00 ©2018 IEEE
Proceeding of the IEEE
International Conference on Information and Automation
Wuyi Mountain, China, August 2018