Published in IET Computer Vision
Received on 1st August 2009
Revised on 27th June 2010
doi: 10.1049/iet-cvi.2009.0075
ISSN 1751-9632
Robust mean-shift tracking with corrected
background-weighted histogram
J. Ning
1,2,3
L. Zhang
2
D. Zhang
2
C. Wu
1
1
The State Key Laboratory of Integrated Service Networks, Xidian University, Xi’an, People’s Republic of China
2
Department of Computing, The Hong Kong Polytechnic University, Kowloon, Hong Kong, People’s Republic of China
3
College of Information Engineering, Northwest A&F University, Yangling, People’s Republic of China
E-mail: cslzhang@comp.polyu.edu.hk
Abstract: The background-weighted histogram (BWH) algorithm proposed by Comaniciu et al. attempts to reduce the
interference of background in target localisation in mean-shift tracking. However, the authors prove that the weights assigned
to pixels in the target candidate region by BWH are proportional to those without background information, that is, BWH does
not introduce any new information because the mean-shift iteration formula is invariant to the scale transformation of
weights. Then a corrected BWH (CBWH) formula is proposed by transforming only the target model but not the target
candidate model. The CBWH scheme can effectively reduce background’s interference in target localisation. The
experimental results show that CBWH can lead to faster convergence and more accurate localisation than the usual target
representation in mean-shift tracking. Even if the target is not well initialised, the proposed algorithm can still robustly track
the object, which is hard to achieve by the conventional target representation.
1 Introduction
Object tracking is an important task in computer vision. Many
algorithms [1] have been proposed to solve the various
problems arisen from noises, clutters and occlusions in the
appearance model of the target to be tracked. Among
various object-tracking methods, the mean-shift tracking
algorithm [2–4] is a popular one because of its simplicity
and efficiency. Mean shift is a non-parametric density
estimator that iteratively computes the nearest mode of a
sample distribution [5]. After it was introduced to the field
of computer vision [6], mean shift has been adopted to
solve various problems, such as image filtering,
segmentation [7–11] and object tracking [2, 3, 12 –18].
In the mean-shift tracking algorithm, the colour histogram
is used to represent the target because of its robustness to
scaling, rotation and partial occlusion [2, 3, 19]. However,
the mean-shift algorithm is prone to local minima when
some of the target features are present in the background.
Therefore Comaniciu et al. [3] further proposed the
background-weighted histogram (BWH) to decrease
background interference in target representation. The
strategy of BWH is to derive a simple representation of the
background features and to use it to select the salient
components from the target model and target candidate
model. More specifically, BWH attempts to decrease the
probability of prominent background features in the target
model and candidate model and thus reduce the
background’s interference in target localisation. Such an
idea is reasonable and intuitive, and some works have been
proposed to follow this idea [20–22].In[20], the object is
partitioned into a number of fragments and then the target
model of each fragment is enhanced by using BWH.
Different from the original BWH transformation, the
weights of background features are derived from the
differences between the fragment and background colours.
In [21], the target is represented by combining BWH and
adaptive kernel density estimation, which extends the
searching range of the mean-shift algorithm. In addition,
Allen et al. [22] proposed a parallel implementation of
mean-shift algorithm with adaptive scale and BWH and
demonstrated the efficiency of their technique in a single
instruction multiple data computer. All the above BWH-
based methods aim to decrease the distraction of
background in target location to enhance mean-shift
tracking. Unfortunately, all of them do not notice that the
BWH transformation formula proposed in [3] is actually
incorrect, which will be proved in this paper.
In this paper, we demonstrate that the BWH algorithm will
simultaneously decrease the probability of prominent
background features in the target model and target
candidate model. Thus, BWH is equivalent to a scale
transformation of the weights obtained by the usual target
representation method in the target candidate region.
Meanwhile, the mean-shift iteration formula is invariant to
the scale transformation of weights. Therefore the mean-
shift tracking with BWH in [3, 20–22] is exactly the same
as the mean-shift tracking with usual target representation.
Based on the mean-shift iteration formula, the key to
effectively exploit the background information is to
decrease the weights of prominent background features. To
this end, we propose to transform only the target model but
62 IET Comput. Vis., 2012, Vol. 6, Iss. 1, pp. 62 –69
&
The Institution of Engineering and Technology 2012 doi: 10.1049/iet-cvi.2009.0075
www.ietdl.org