Robust visual tracking via self-similarity learning

Self-similarity is an attractive image property that has been successfully applied to object recognition due to its robustness to severe target appearance variations. However, less attention has been paid to explore self-similarity for visual tracking, mainly because it is difficult to learn self-similarity information between different features suitable for visual tracking. To address this issue, a simple, yet effective approach is presented to learn self-similarity information among the local features extracted from the different regions of the target. The target is first divided into some non-overlapping regions, in which each region is described by the histogram of gradient (HOG) features. Then, an explicit polynomial kernel feature map is constructed, which is capable of characterising the self-similarity information among all the local regions in the targets. Finally, based on the feature maps, a linear support vector machine (SVM) is learnt via an online dual coordinate descent method that offers fast convergence guarantee. Experiments on a large tracking benchmark dataset with 50 sequences demonstrate the superiority of the proposed method over state-of-the-art methods.