Robust Dual-Kernel Tracking Using Both Foreground and Background - Advances in Image and Graphics Technologies - page 83

Robust Dual-Kernel Tracking Using Both Foreground
and Background

Wangsheng Yu, Zhiqiang Hou, Xiaohua Tian, Lang Zhang, and Wanjun Xu

Information and Navigation College, Air Force Engineering University, Xi'an, China
[e-mail]

Abstract. The kernel-based mean shift tracker outperforms other trackers due
to its innovated target representation and efficient optimization strategy. How-
ever, this representation relies overmuch on the foreground and thus, decreases
the robustness to the background change and clutter. To this point, this paper
presents a dual-kernel tracker based on mean shift using both foreground and
background. The proposed target representation consists of foreground model
and background model, and the optimizing process integrates foreground kernel
iteration and background kernel iteration. Experiments indicate that the pro-
posed tracker obtains better performance in coping with background change and
clutter.

Keywords: Visual tracking, mean shift, kernel-based tracker, dual-kernel
tracker.

1 Introduction

Visual tracking is widely used in civil and military fields. As one of the famous track-
ers, the kernel-based tracker [1] is essentially a model-driven tracker based on an effi-
cient searching algorithm. It searches the local maxima along with the ascent direction
of gradient in feature space [2], which is known as mean shift iteration. Due to its
simplicity and efficiency, mean shift algorithm has been widely applied in visual
tracking, image smooth, cluttering and segmentation [3]. Another technique presented
in kernel-based tracker is target representation. It spatially masks the target window
with an isotropic kernel and transforms it into histogram features. In the past decade,
many researchers did further studies to improve the tracking performance of kernel-
based tracker. The fruitful works varies from kernel bandwidth selection [4], spatial
histogram [5], adaptive binning histogram [6], anisotropic kernel [7], background
contrasting [8], multi-part model [9], to more discriminative similarity metric [10].
The common point shared among these works is that a single kernel and limited
background information is utilized.
In this paper, we proposed a dual-kernel tracker using both foreground and back-
ground, that is, both foreground kernel and background kernel are utilized to accom-
plish the tracking task. The fully utilizing of sufficient background information
obviously improved the tracking accuracy and robustness.