Improved kernelized correlation filters tracking algorithm with adaptive learning factor

摘要

Tracking with kernelized correlation filters is a new idea recently proposed which is different from traditional methods based on target features. This method achieves fast tracking speed, however, it is seriously compromised when the tracking target has large-scale changes and severe occlusion. An improved update model based on kernelized correlation filters is proposed in this paper to effectively overcome the above problems. An adaptive learning factor is defined with Peak-to-sidelobe ratio which estimates the correlation between different candidate images. It achieves adaptive online update of the tracking model. Experiments demonstrates that the presented algorithm can adjust the learning factor in real time according to different scenarios, which results increased success rate of tracking. With the adaptive learning factor, the presented algorithm shows advanced adaptability to partial occlusions, illumination, and target scale variations.