Relaxed Target Representation for Visual Tracking

摘要

In visual tracking, developing a robust appearance model is a challenging issue due to the influences of occlusion, drastic illumination variation, background clutters and rotation. Existing appearance tracking algorithms represent a target candidate by a linear combination of a set of dictionary templates. The feature elements of a target candidate have the same coding vector. However, due to influences of appearance variations and different spatial locations of pixels, different feature elements have different importance. So the coding vector coefficients of dictionary templates should have some diversity to reflect the distinctive importance of feature elements. In this paper, we propose a novel appearance model that considers the similarity and distinctiveness of feature elements. A weighted regularized term is introduced in representing a target candidate. Remarkable performance on challenging sequences demonstrate the effectiveness of the appearance model and the robustness of the proposed tracker.