Self-tuning Motion Model for Visual Tracking

In visual tracking, how to select a suitable motion model is an important problem to deal with, since the movements in real world are always irregular in most cases. We propose a self-tuning motion model for target tracking in this paper, where the current motion model is computed according to the relative distance of the target positions in the last two frames. Our method has achieved excellent performance when experimenting on the sequences where the targets move unstably, abruptly or even when partial occlusion exists, and the method is particularly robust to the unsuitable initial motion model.