Online Visual Tracking via Structural Compressive Sensing

The strong theoretical support from compressive sensing motivates many researchers to develop various algorithms and nowadays compressive tracking is extremely popular in the visual tracking community. In this paper, a novel structural compressive tracker is proposed. The contributions compared with traditional compressive trackers can be summarized into three aspects. First, the motion information is effectively integrated into compressive sensing based appearance model by introducing particle filter motion estimator. Second, a 3-order transition model is designed to simultaneously consider the velocity and acceleration to estimate both the location and scale of the target object. Third, the structural holistic appearance information is efficiently embedded into our observation model, which further provides additional constraints to avoid potential drift. Extensive experiments on several benchmark sequences demonstrate the favored performances of the proposed method.