Modified Kernelized Correlation Filter Tracker Based on Saliency Detection and Reliability Judgment

Abstract

With the rapid development of the correlation filter and deep learning technology, object tracking has been applied widely in the field of autonomous driving and video surveillance. It is challenging to achieve robust and efficient object tracking due to the variability of scenes and the complexity of the background. Considering real-time and computation limitations, correlation filter based tracker is still a good solution. In this paper, we proposed a modified tracking algorithm based on kernelized correlation filter. To distinguish the object from the background noise and eliminate unreasonably high energy values of the correlation filter in the boundary region, the saliency detection of object candidate regions is adopted. Besides, a compound discrimination method is proposed considering both the maximum correlation peak and the average peak-to-correlation energy to judge the reliability of tracking results accurately. Our approach is evaluated on OTB-2015 dataset and the experimental results show that our approach achieves outstanding performance than the classical algorithm KCF. Moreover, it is robust and efficient enough for occlusion and illumination change.