A combined visual tracker based on global appearance and local features

摘要

Visual tracking is a vital task of computer vision, and becoming the basis of automated video surveillance. In this paper, we propose a novel tracking method which integrates two complementary trackers together, tracker A based on global appearance and tracker B based on a dynamic set of local features of the tracked object. Tracker A, an enhanced mean-shift tracker using the posterior probability measure, could somehow decrease the interference of background pixels enclosed in object model. Tracker B could obtain motion estimation of the tracked object from local feature matching using SURF-RANSAC method. The combination strategy largely expands the range of applications by changing the representation of the tracked object and tracking method flexibly. Experimental results on the dataset in VOT Challenges demonstrate robustness and accuracy of the tracker.