The Multiple Model Vo-Vo Filter
Wei Yi, Meng Jiang, Reza Hoseinnezhad
Abstract—A Jump Markov System (JMS) multiple model
solution is presented for tracking highly dynamic targets using
the Vo-Vo filter (also known as Generalized Labeled Multi-
Bernoulli or GLMB filter). The closed-form solution is derived for
Gaussian mixture implementation of the multiple-model Vo-Vo
filter. The performance of the proposed method is examined and
compared with the state of art in challenging scenarios involving
numerous appearing and disappearing targets with randomly
time-varying dynamics. The results demonstrate that in such
applications, the proposed method can outperform the competing
techniques in terms of tracking accuracy, and is highly robust
to variations in application and filter parameters such as clutter
rate and model transition probabilities in the JMS.
I. INTRODUCTION
B
AYESIAN filters, in general, are the most popular type
of solutions developed and used for multi-target tracking
problems. The most commonly used Bayesian multi-target
filters include JPDA [1], MHT [2], PHD [3], CPHD [4], Multi-
Bernoulli [3], [5], and Labeled Multi-Bernoulli (LMB) [6]
filters, and most recently the Vo-Vo filter (also called Gener-
alized Labeled Multi-Bernoulli or GLMB filter) [7], [8].
1
The
main advantage of the above filters is the neat and efficient
machinery they provide to harvest all the information that is
available in relation to the trackable targets. In general, the
application-specific information regarding the target dynamics,
birth and survival are used in the prediction step, then the
measurements (normally acquired by sensors) are processed
in the update step of the filter.
In their standard form, the above multi-target filtering
solutions are implemented based on the assumption that the
states of all targets follow the same stochastic dynamic model
at all times. Examples of such models include Brownian
motion models, nearly constant velocity (NCV) models, nearly
constant acceleration (NCA) models and coordinated turn (CT)
models [10].
In many real world applications, assuming that all targets
evolve according to the same dynamic model at all times, is
far from accurate. For example, in traffic monitoring, the dy-
namics of some slower targets may fit well to an NCV model
while faster maneuvering targets may follow an NCA model
more closely. In battlefield surveillance applications with radar
tracking, a maneuvering target may fly very differently at
different times, hence its dynamics would match different
models in different periods of time. In such applications, the
W. Yi and M. Jiang are with University of Electronic Science and Technolo-
gy of China, Chengdu, Sichuan, P.R. China, 611731. Fax: +86-028-61830064,
Tel: +86-028-61830768, E-mail: kussoyi@gmail.com.
R. Hoseinnezhad is with the School of Engineering, RMIT University,
Victoria 3083, Australia, Tel: +61-3-9925-6135, E-mail: rezah@rmit.edu.au.
This work was supported by the Australian Research Council’s Discovery
Project Program, via the ARC Discovery Project grants DP130104404 and
DP160104662, and supported by the National Natural Science Foundation of
China under Grants 61301266, the Chinese Postdoctoral Science Foundation
under Grant 2014M550465 and Specail Grant 2016T90845.
1
We follow Mahler who used the name “Vo-Vo filter” in his book [9].
standard implementation of the aforementioned filters (with
a fixed motion model) may not be able to accurately track
multiple rapidly maneuvering targets. Thus, new Bayesian
multi-target filters that consider multiple models characterizing
different maneuvers are necessary to describe the motion of
maneuvering targets.
To incorporate multiple models into existing Bayesian
multi-target filtering schemes, the popular approaches in the
literature are the Jump Markov System (JMS) approach [11]
and its practical approximation, namely, the interacting mul-
tiple model (IMM) algorithm [12]. In the IMM algorithm, a
mixing step is added to the filter (in addition to prediction and
update steps). While the original solution [12] was developed
for filtering of linear systems with Markovian switching coef-
ficients (dynamic multiple model systems), the IMM approach
to hypotheses merging was used to devise multiple-model
extensions of the JPDA [13] and MHT [14] filters. In JMS
approach, the target state is augmented with an additional
motion model label, and the augmented state of each target
evolves with time according to a finite state Markov chain [11].
This approach has been employed to formulate multiple-model
extensions of PHD [15], GM-PHD [16], CPHD [17], multi-
Bernoulli [18] and LMB [19] filters.
This paper incorporates multiple models into Vo-Vo fil-
ter [7], [8] and gives the specific closed-form solution for
multiple-model extension, called MM-Vo-Vo filter, based on
the JMS approach. The Vo-Vo filter uses a special class of
random finite set (RFS) distribution called generalized labeled
multi-Bernoulli (GLMB). Multi-target Bayesian filtering of
GLMB random finite sets enables the estimation of target
tracks as well as their states.
The GLMB prior is closed under the Chapman-Kolmogorov
equation, and more importantly, it is a conjugate prior for the
point measurement set likelihood model. Vo-Vo filter is one
of the few stochastic multi-target filters in the literature that
satisfy this conjugacy property. This important characteristic
and the efficient implementation of the Vo-Vo filter presented
in [8], have led to its outstanding performance in terms of both
tracking error and computational cost. As a result, the Vo-
Vo filter has been increasingly adopted in many applications
from track-before-detect [20], to tracking with superpositional
sensors [21], to visual tracking [22]. This paper addresses an
important gap in the literature in relation to using Vo-Vo filter
to track maneuvering targets with dynamics that can follow
multiple models [23], [24]
2
.
We introduce specific recursion formulae for Vo-Vo filtering
of augmented target states under multiple motion models.
MM-Vo-Vo filter uses JMS to extend the standard Vo-Vo
recursion to allow for multiple target motion models. This
extension is implemented using Gaussian mixture approxi-
mations. The multiple model implementation is validated a-
gainst its respective standard implementations of PHD, CPHD,
LMB and Vo-Vo filters in a challenging scenario involving
up to ten appearing and disappearing targets that change
their motion behaviour at different random times. The OSPA
2
These two papers have independently proposed methods for maneuvering
target tracking based on Vo-Vo filter.