Robust visual tracking via guided low-rank subspace learning

Subspace methods have attracted increasing attention for visual tracking. However, most previous work only aim to pursuit the subspace basis to represent appearances, thus cannot reveal the rich structure information in real world videos. This paper proposes a guided low-rank subspace learning framework to simultaneously extract the orthogonal subspace basis, the low-rank coefficients and the sparse errors to build observation model. Benefiting from the predefined guidance, we can successfully extract the relationship between the candidate particles and the subspace basis, thus most of the background can be suppressed. By reformulating the proposed model into two simple subproblems, we further develop an efficient online optimization scheme for our tracking system. Extensive experiments well validate the effectiveness and stability of our tracker over other state-of-the-art methods.