Visual Tracking with Multi-level Dictionary Learning

Abstract

In this paper we propose a robust sparse based visual tracking method by exploiting local representations in a particle filter framework. We construct a Multi-level Local Dictionary which consists of positive templates and negative templates for discriminative model, Which divide the positive and negative dictionary into two levels called static templates and dynamic templates, respectively, thus can account for the targets appearance changes. An effective method is also introduced which calculate the confidence value with takeing local information into consideration, which makes the value more accurate. Furthermore, an online dictionary learning algorithm is proposed. We update the dynamic positive and negative templates separately. Specially, update the negative templates more frequently, but keep the static templates constant. We test our proposed method on several challenging video sequences and numerous experiments had proved it can performs an excellent results comparing to several state-of-the-art algorithms for it can deal with appearance changes and occlusion effectively and efficiently.