Object tracking based on local learning

In this paper, a novel object tracking algorithm based on local learning is proposed. We train a feature-based distance function as a local model for each training sample by using local learning method, which has been shown to be effective to tackle large intra class variations. In the tracking process, distances between testing and training samples are obtained by the trained distance functions, and then object tracking is accomplished by searching for the candidate with smallest weighted sum of distances from all positive training samples. Experimental results demonstrate that the proposed tracking algorithm based on local learning is robust in handling occlusion, motion blur, and rotation, which are prone to cause intra class variations.