JFT: A Robust Visual Tracker Based on Jitter Factor and Global Registration

Haopeng Wang, Shuixing Pan, Dilong Li, Yueqiang Zhang, Linyu Huang, Hongxi Guo, Xiaolin Liu, Qifeng Yu

Abstract: Visual object tracking is an important yet challenging task in computer vision, whose accuracy is highly subject to the problems of camera motion/shake and occlusion. In order to solve these two challenging problems and improve the tracking accuracy of real scenes, this paper proposed a robust visual object tracker based on Jitter Factor and global registration. The proposed tracker firstly extracts the histogram of oriented gradient (HOG) features and color features of the target object to train the correlation filter. When the response map is unreliable, the proposed tracker treats the tracking problems as a global background motion and target motion problem, and then evaluates the state (tracking or missing) of the target by using Jitter Factor. If the target is assumed to be missing, global image registration and correlated Kalman filter would be applied to correct and predict the corrected target position. Experimental results on RGB-T234 show that after introducing the proposed Jitter Factor and global image registration, the correlation-filter-based trackers gained a ≥ 2.2% increase in precision rate.