Multilayer feature combination for visual tracking

This paper proposes a new tracking method based on multilayer feature combination. In each layer, the target is segmented into local patches, and patch sizes of different layers are different. Through this way, each layer contains different local information of the object, which are mutually complementary for each other. For each layer, the local patches are represented with sparse codes. We combine these sparse codes into a histogram of sparse codes (HSC) for each layer. To handle appearance variations, we improve the HSC and obtain a modified histogram of sparse codes (MHSC), which is used to represent each layer. We combine the MHSCs of multilayer to form the feature vector of the object. To improve the robustness of the feature vector, different weights are assigned to various layers because each layer has different discriminative power under different cases. Within Bayesian framework, we achieve visual tracking by finding the candidate which has the highest similarity with the reference. Experiments demonstrate that the proposed method outperforms several state-of-the-art trackers.