Self-tuning visual tracking based on multiple motion models

摘要

To deal with irregular movements in visual tracking, such as abrupt motion in velocity and direction, we propose a self-tuning tracking method based on multiple motion models (SBMM). Multiple motion models are qualitatively defined as high-speed, middle-speed and low-speed motion models and particle filter is utilized to realize the algorithm. In our algorithm, particles are divided into multiple groups to employ different motion models mentioned above. Furthermore, according to the circumstances of the motion in previous frames, we dynamically adjust the number of particles in each group during the entire tracking process to make the tracker more robust. Our method achieves excellent performance in experimental sequences where objects move unstably, abruptly and even in partial occlusion.