Multiscale salient region-based visual tracking

Abstract

This paper proposes a novel visual model to detect the salient regions of the target in complex tracking scenarios. The main idea of the proposed visual model is to generate an overcomplete set of local image patches to describe the multiscale regions of the target, and select the most important and reliable regions. The importance of each patch is evaluated by its stability and discrimination in the local feature space, while the reliability is measured by the contrast of the target and its surrounding background in the global feature space. By combining the importance and reliability, the salient regions are selected from the patch set to represent the target. Experimental results on benchmark video sequences show that the proposed visual model can improve the tracking performance effectively.

http://ift.tt/2oJXYVD