Kernelized Target Representation for Visual Tracking with Sparse Constraint

Abstract

Robust visual tracking is a challenging task due to factors motion blur, fast motion, partial occlusion and illumination variation. Existing tracking algorithms represent a target candidate by templates or a linear combination of them with some constraints such as sparse coding. While the high computational cost restricts the tracking speed with sparse constraint since many trivial templates are introduced. Due to affecting by complicated appearance variations, the relationship between a target candidate and the corresponding target templates is not linear. Maybe it is nonlinear or complex. In this paper, we present a kernelized target representation model with sparsity constraint for visual tracking. Namely, a target candidate is represented by a nonlinear combination of templates with sparsity constraint. The proposed appearance model have the advantages of sparse coding and kernel method, which is robust to outliers and represents a target candidate in a high feature space. A novel tracker is proposed upon the presented appearance model. Superior experimental results are achieved against state-of-the-art trackers on some challenging sequences.