Structured Sparse Representation Visual Tracking Using Bayes Classifier

In this paper we propose a structured sparse representation based visual tracking algorithm by using both the generative appearance model and the discriminative model. Firstly, structured sparse representation models are used to exploit both holistic and local information of the target. In the structured sparse representation framework, a new over-complete dictionary containing both target and background templates is proposed to enhance the robustness of the tracking algorithm. Secondly, the tracking task is treated as a binary classification problem, and a Bayes classifier is trained online by using structured sparse codes of positive and negative samples. Furthermore, a kind of residual error score is constructed to improve the detective ability of the tracker. Finally, target templates are updated via a strategy which combines incremental subspace learning and sparse representation, and background templates are updated by samples around latest results. Compared with 4 state-of-the-art tracking algorithms in 6 challenging video sequences, the proposed tracking algorithm demonstrates better performance than other algorithms in terms of experimental results.