An Improved Kernelized Correlation Filter with Redetection Mechanism for Visual Tracking

Abstract

The correlation filters are the core components of most trackers which achieve the excellent results both on accuracy and real-time. However, the previous trackers are prone to drift away from the target when dealing with the challenging situations, e.g. motion blur and fast motion, which leads to the performance of trackers degration. To solve the above problems, in this paper, we propose a novel algorithm which adds the re-detection mechanism to the traditional kernelized correlation filter for judging and verifying the reliability of the detected target before updating the model. We employ the average peak-to-correlation energy to evaluate the confidence level of the candidate position. The experiment results show that proposed algorithm is more accurate and successful than the traditional algorithms.