Object tracking via fragment-based multi-task sparse state inference

Object tracking is an important issue in computer vision and has many potential applications. This paper cast the tracking problem as a sparse representation problem, in which the tracked object is sparsely represented by a series of candidate samples in each frame. For both object template and candidate samples, their observation image patches are divided into multiple fragments to model the feature and spatial information at the same time. Then the state inference processing can be viewed as a multi-task learning problem, which can be solved by the accelerated proximal gradient (APG) method. Finally, we design a generative tracker based on the proposed model and a simple online update manner. To evaluate our tracker and compare it with other popular tracking algorithms, we conduct several experiments on some challenging image sequences. Both qualitative and quantitative evaluations illustrate that our tracker achieves better performance than other trackers.