Multi-target tracking by enhancing the kernelised correlation filter-based tracker

Abstract

A new tracking method based on the kernelised correlation filter (KCF) method is proposed. The tracker improves KCF-based trackers by adding seven proposed components, namely, the motion model, background subtraction, occlusion handling, hijacking handling, object proposal, bounding box modification, and object re-detection. With these components, the tracker robustly tracks multiple targets despite severe occlusion, rapid motion, and the presence of other objects with similar appearance. The visual tracking performance is evaluated by using challenging basketball game videos. Experiments demonstrate that the tracker outperforms the original KCF tracker and other state-of-the-art tracking methods.