Similarity Fusion for Visual Tracking

Multiple features' integration and context structure of unlabeled data have proven their effectiveness in enhancing similarity measures in many applications of computer vision. However, in similarity based object tracking, integration of multiple features has been rarely studied. In contrast to conventional tracking approaches that utilize pairwise similarity for template matching, our approach contributes in two different aspects. First, multiple features are integrated into a unified similarity to enhance the discriminative ability of similarity measurements. Second, the neighborhood context of the samples in forthcoming frame are employed to further improve the measurements. We utilize a diffusion process on a tensor product graph to achieve these goals. The obtained approach is validated on numerous challenging video sequences, and the experimental results demonstrate that it outperforms state-of-the-art t racking methods.