Dual-template adaptive correlation filter for real-time object tracking

Abstract

Visual object tracking is a hot topic in the field of computer vision. The drift of the tracking box or loss of the tracking target often occurs for existing correlation filter based trackers when the target moves quickly or deforms. Focusing on this problem, we propose a dual-template adaptive correlation filter for real-time object tracking. First, we trained templates for different size levels. Second, the best template was selected based on the target response confidence during estimation of the target translation. Third, the dual templates, scale estimation component, and feature fusion component were integrated into the benchmark tracker, the kernelized correlation filter. The object tracking benchmark was used to evaluate the performance of the proposed algorithm. The experimental results show that compared with the benchmark tracker, the average overlap precision and distance precision of this proposed algorithm are increased by 23.2% and 9.4% in OTB-100. The average running frame rate reaches 42 frames per second, which can meet the real-time requirements. At the same time, five algorithms, DSST, SAMF, KCF, CN, and CSK, appear to drift or even lose the target among the four selected typical video sequences, while our algorithm can successfully track the target.