Occlusion handling approach in visual tracking based on multiple-kernel fusion

A novel visual tracking approach based on multiple-kernel fusion was proposed to improve robustness and accuracy of tracking under large-area occlusion. Unlike traditional single symmetric kernel weighted histogram used in mean shift tracking, this approach adopted several asymmetric kernel functions centered at different positions within target region to build a set of asymmetric kernel weighted histograms. Because these histograms weighted each part of the target region differently, there must be some less influenced histograms during occlusion. Based on each histogram, a set of target location estimations were provided respectively by mean shift iteration, and the target location was obtained by fusing these estimations using Dempster-Shafer evidence theory. The experimental results demonstrate the effectiveness of the proposed approach under large-area occlusion.