Frequency decomposition-based tracking for improving a CSK tracker

Abstract

Object tracking is an important task within the field of computer vision. Object tracking methodology is divided into three categories: point-, kernel-; and silhouette-based tracking. Recently, a correlation-based kernel tracking has been used in object tracking with high accuracy and performance. However, the correlation-based approach heavily relies on luminance information, thus it has many problems. In addition, various environments including illumination variation, deformation, occlusion and scale variation make the object tracking unreliable especially in chronic problem such as state update. To isolate this problem, we present a frequency decomposition-based effective approach for more robust circulant structure kernel (CSK) traker. Experimental results show that the proposed approach outperforms the conventional CSK tracker by about 5% on average for 38 illumination changing sequences from the commonly used 100 test sequences.