Online Tracking via Superpixel and Sparse Representation

Handling appearance variations is a very challenging issue for visual tracking. In this paper, an effective superpixel based L1 tracking method (SuperPixel-L1 tracker, SPL1) is proposed to deal with the above problem in a particle filter framework. First, the mid-level visual cue with structural information is exploited to construct the dictionary and model the object appearance. Then each candidate state defined by a particle is solved via L1 minimization. The candidate with the smallest reconstruction error is selected as the tracking result. Finally, the online dictionary updating strategy is further improved. The dictionary needs to be updated regardless of whether the object is occluded or not. The initial frame information is retained during the updating process to reduce the possibility of the object drift. Simulation results show that SPL1 tracker can still stably track the object under the circumstance of long-term occlusion, large scale and illumination changes.