A visual tracker based on improved kernel correlation filter

Abstract

Recently, Kernel Correlation Filter (KCF) is introduced into visual tracking, which has shown to provide the highest speed among the top ten trackers. However, the KCF tracker could not perfectly handle the occlusion during target's moving. Leading to the failure of tracking. A robust occlusion detection scheme was proposed based on the KCF framework. When the object was under partially and fully occlusion, the kalman filter was used to predict the state information and stop the classifier model to update at the same time, which improved the ability of occlusion handling. The experimental results on comparison with some state-of-the-art trackers such as Struck, KCF, TLD and MIL demonstrate that the proposed method could deal with the occlusion without any additional complex computing consumption.