An Improved Kernel Correlation Filter for Occlusion Target Tracking

In order to deal with the issues of stable tracking of moving targets under the occlusion situation, an improved kernel correlation filter is proposed in this paper. Color features can accommodate deformation, but can’t handle changes in illumination or the situation that the background is similar to the target. The local areas of the image are used to construct the target features in the histogram of oriented gradient (HOG), which is not easily affected by the illumination changes, but can’t adapt to the rapid deformation of the target. Therefore, the fusion features of HOG and color name (CN) are used in this paper. Due to the lack of occlusion discrimination mechanism, the cyclic structure in kernel correlation filter (KCF) is easy to introduce contaminated occluded samples, resulting in tracking failure. In this paper, the peak to sidelobe ratio (PSR) value of each frame is obtained and compared with the set threshold. When the occlusion is determined, a strategy for predicting the trajectory of the occluded target using a Kalman filter with the history frames is adopted. The proposed method is superior to KCF、 CSK、SCM、Struck、TLD and CT algorithms on the OTB-2013 benchmark dataset, especially under the occlusion situation.