Long-term tracking algorithm of soccer player based on kernelized correlation filter

Abstract

The performance of player object tracking in soccer match video is seriously affected by challenges like occlusion, out-of-view, similarity interference, low resolution, etc. To solve this problem, we propose a long-term tracking algorithm based on kernelized correlation filter. Firstly, the algorithm fuses shape, color and grayscale features to enhance the object representation ability and introduces scale filter to realize real-time scale estimation to improve the tracking accuracy and robustness. Secondly, the algorithm monitors tracking status in real time by the tracking result evaluation function. If the status judged good, the multi-peak re-detection of response map is used to review the tracking result. Otherwise, the object is re-detected through sliding windows to realize long-term tracking. The experimental results tested on Benchmark for Soccer Player Tracking (BSPT) demonstrate that the proposed tracker can achieve an accurate, real-time and long-term visual tracking for soccer player while runs at speed near 80 FPS.