Object tracking based on local dynamic sparse model

• Local patches of a tracked object are represented by their dictionary updated online.
• The inter-frame correlation between sparse representations of patches is modeled.
• The dependency of sparse coefficients between patches in a frame is modeled.
• The effect of occluded patches is eliminated in the update of templates.

Sparse representation has been widely applied in many objecting tracking methods. In this paper, we present a robust and effective object tracking approach based on the local dynamic sparse model, called Local Dynamic Sparse Tracking (LDST). In the proposed method, the local patches of a tracked object are linearly represented by their respective dictionary updated online, and the inter-frame correlation between sparse representations of corresponding patches are modeled in the time domain. To further improve its robustness, the dependency of sparse coefficients between patches in each frame is also characterized by the ℓ1,2ℓ1,2 mixed norms. In addition, for each patch, different weights are exploited in calculating the likelihood probability, in order to eliminate the effect of occluded patches when updating templates. The evaluation experiments on the challenging sequences demonstrate that the proposed method has the better performance compared with some typical state-of-the-art methods.