Robust visual tracking with discriminative sparse learning

摘要

Recently, sparse representation in the task of visual tracking has been obtained increasing attention and many algorithms are proposed based on it. In these algorithms for visual tracking, each candidate target is sparsely represented by a set of target templates. However, these algorithms fail to consider the structural information of the space of the target templates, i.e., target template set. In this paper, we propose an algorithm named non-local self-similarity (NLSS) based sparse coding algorithm (NLSSC) to learn the sparse representations, which considers the geometrical structure of the set of target candidates. By using non-local self-similarity (NLSS) as a smooth operator, the proposed method can turn the tracking into sparse representations problems, in which the information of the set of target candidates is exploited. Extensive experimental results on visual tracking have demonstrated the effectiveness of the proposed algorithm.