Online learning of region confidences for object tracking

This paper presents an online learning method for object tracking. Motivated by the attention shifting among local regions of a human vision system during tracking, we propose to allow different regions of an object to have different confidences. The confidence of each region is learned online to reflect the discriminative power of the region in feature space and the probability of occlusion. The distribution of region confidences is employed to guide a tracking algorithm to find correspondences in adjacent frames of video images. Only high confidence regions are tracked instead of the entire object. We demonstrate feasibility of the proposed method in video surveillance applications. The method can be combined with many other existing tracking systems to enhance robustness of these systems.