A KCF-Based Tracker with Faster Scale Estimation and Effective Occlusion Detection

Abstract

In visual object tracking, designing a robust and fast tracking algorithm is always a challenging task. KCF is exactly one of those trackers which have fast computational speed and well tracking accuracy. However, when the object’s scale variates, the tracking accuracy of KCF will greatly decrease due to the lack of scale estimation approach. For this reason, DSST has proposed a KCF-based robust scale estimation approach to tackle the problem of scale variation. But in order to meet the need of practical application, computational efficiency is always the highest priority of a tracking algorithm. Besides, partial occlusion often happens in an object tracking mission and significantly affect the tracking performances in the real situations. In this paper, an approach that estimate the scale variation faster is proposed based on KCF and DSST. Meanwhile, an approach which is designed for effective occlusion detection is also proposed, which improves the tracking accuracy of KCF in the situations of object occlusion. Finally, the improved tracker based on KCF and DSST has about 7% increase on tracking accuracy and perform faster in most cases. It turns out that the proposed approaches perform better on scale estimation and occlusion detection in the real situations.