A robust local sparse tracker with global consistency constraint

• The first contribution is proposing a novel local sparse model with global constraint, which can reduce the drifting problems.
• The second contribution is giving a two-stage algorithm to exploit the partial and spatial information without overfitting problems.
• The third contribution is raising exploiting occlusion information to guide the template update, which can improve the template update strategy.

In the field of visual object tracking, partial occlusion and the variation of illumination, pose and background are the core problems to be handled. More and more visual tracking methods tend to exploit part or local features to deal with the above problems. However, single local features may lead to overfitting and drifting problem, as will cause the failure of tracking task. In this paper, we propose a novel tracking method by exploiting the partial and spatial information with a global regulation on the stabilization of local features. With the local features and the global constraint, the problems of occlusion and variation can be well solved and a stable performance can be obtained without overfitting. In the first stage, overlapped patches are used to hold the local features and each patch is reconstructed with all the template patches. The reconstruction coefficients are obtained by solving the ℓ1 regularized least square problem. In the second stage, a global constraint is added to find the final result. The constraint is achieved by restraining the difference in contributions of each patch. Additionally, we employ occlusion information to improve the template update strategy. The experiment results on several widely used benchmark datasets demonstrate that our method is effective and outperforms the state-of-the-art trackers.