GPU-Accelerated Video Background Subtraction Using Gabor Detector
q
Lixia Qin
a
, Bin Sheng
a,b,
⇑
, Weiyao Lin
a
, Wen Wu
c
, Ruimin Shen
a
a
Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China
b
State Key Lab. of Computer Science, Inst. of Software, Chinese Academy of Sciences, Beijing, China
c
Dept. of Computer and Information Science, Faculty of Science and Technology, University of Macau, Macao, China
article info
Article history:
Received 24 October 2014
Accepted 9 July 2015
Available online 15 July 2015
Keywords:
Background subtraction
ViBe
Background model
Gabor filter bank
Ghost suppression
GPU parallel optimization
Boundary information
Anti-noise
abstract
Background subtraction is a technique in which a background model is built and compared with the cur-
rent frame to distinguish the foreground from the background. The technique is extensively used to facil-
itate automatic detection, segmentation, and tracking of objects in videos. However, conventional
background subtraction methods have disadvantages, such as slow model-updating speeds, the inability
to leverage edge information, and negative anti-noise properties in conditions with illumination varia-
tions. We therefore propose a ViBe-based method that employs simplified Gabor wavelets to calculate
image edge information. The method randomly applies relevant pixels to initialize or update the back-
ground model and considers the variation dispersion degree during segmentation. Experimental results
indicate that the proposed method performs well in foreground–background segmentation and in color
shift situations caused by illumination or aperture adjustments. Moreover, the processing speed of the
proposed approach is accelerated by parallel computing capacity of graphics processing units.
Ó 2015 Elsevier Inc. All rights reserved.
1. Introduction
In video processing, foreground and background segregation is
a key technique. Its main objective is to extract the actual moving
foreground from various video sequences disrupted by noise. This
technique is extensively used to facilitate recognition, segmenta-
tion, and tracking in video technologies. A simple method of fore-
ground and background segregation involves evaluating and
inspecting regions with differences in two consecutive frames to
determine the foreground regions [1]. However, this method
causes deformation of the foreground on account of slow move-
ment; moreover, because it is weakly resistant to ambient noise,
a small amount of noise can significantly affect the segregation.
In practical settings, a significant amount of noise, as well as back-
ground variations, typically exists. Consequently, a significant
amount of current research focuses on foreground and background
segregation of movement in complex circumstances that are
affected by ambient environments or changing backgrounds.
In 2009, Barnich and Van Droogenbroeck [2] presented the ViBe
background modeling algorithm, which is renowned for its rapid
processing and high recognition accuracy. Nevertheless, ViBe has
limitations, such as the issue of blinking dots that are easily
perceived in regions with substantial color differences (e.g., object
edges). This phenomenon occurs because the algorithm selects the
adjacent points with color differences as sample points in the back-
ground model. Owing to the randomized expanding and updating
strategy of ViBe, improvement of the randomized expanding and
updating speed to enhance the ghost subtraction speed causes
the static or slow moving foreground to more quickly blend into
the background. This can result in the foreground being over-
looked. In addition, ViBe shows weak performance in circum-
stances in which the illumination varies.
In this paper, we propose a ViBe-based method that employs
simplified Gabor wavelets to calculate image edge information.
In cases in which the background object is moved, both the new
and original locations of the object are detected as part of the fore-
ground. However, the original place is not a virtual foreground; the
artificial foreground must be suppressed. In our method, edge
information is considered for obtaining a more effective ghost sup-
pression result. If no edge exists between the background and new
foreground, it is assumed to be an artificial foreground and the
speed of the ghost subtraction process is increased. We improve
the performance on illumination variation conditions with a pro-
posed foreground segregation strategy. In addition, we use GPU
parallel optimization to boost the method.
The remainder of this paper is structured as follows. In
Section 2, we review existing researches on background extraction.
The advantages and limitations of each method are presented. In
http://dx.doi.org/10.1016/j.jvcir.2015.07.010
1047-3203/Ó 2015 Elsevier Inc. All rights reserved.
q
This paper has been recommended for acceptance by M.T. Sun.
⇑
Corresponding author at: Department of Computer Science and Engineering,
Shanghai Jiao Tong University, Shanghai, China.
J. Vis. Commun. Image R. 32 (2015) 1–9
Contents lists available at ScienceDirect
J. Vis. Commun. Image R.
journal homepage: www.elsevier.com/locate/jvci