Visual Tracking via Local Sparse Correlation Filters

Visual tracking is a challenging problem due to the intricate appearance variation of the objects in video sequences. Recently, correlation filters(CFs) technique has become a powerful tool for building a robust and high-speed visual tracker. However, there are still some intractable problems need to be solved: 1) The updating strategy of the CF's appearance model is linear, this strategy can not distinguish objects from the occlusions, may adding non-objects to the linear appearance model; 2) The conventional CFs can not handle the affine transforms of the objects. In this paper, we combine the local sparse method and CFs to construct an appearance model of the objects, and use the particle filters to find the objects' affine transforms. The experiments show that our approach outperforms the original local sparse coding approach and other state-of-the-art trackers.