Covariance based local salient descriptors for visual tracking

When visual tracking is performed by human, we typically pay attention to some salient regions or points of the target instead of the whole target. Inspired by this visual saliency property of human visual system, the paper proposes a novel salient regions extraction method to model target appearance. In order to capture the salient and spatial information within this model, the method extracts a set of local salient descriptors based on covariance features from the target. Afterwards, an optimization problem is constructed with respect to the features of these salient regions, and the optimal target state is obtained by solving this problem using a gradient descent algorithm. Experiments on several challenging video sequences demonstrate the good performance of the proposed method compared with four state-of-art tracking methods.