A Target Tracking Method Based on KCF for Omnidirectional Vision

Omnidirectional vision can solve the problem that the tracking object easily lost in perspective vision with narrow sight.In this paper, we propose a tracking method based on Kernel Correlation Filtering(KCF) for omnidirectional vision.For the insufficiency of the KCF algorithm, we make some improvements about scale and occlusion.In our algorithm, we combine the multi-scale KCF with kalman estimate. In addition, occlusion judgement and re-detection are added to our algorithm. Due to the large filed-of-view and distortion of the omnidirectional, the target is re-detected using SURF in a local area around the predictive window. Experiments show that our method has better long-term tracking performance in the case of scale variation, occlusion for omnidirectional vision system.