A visual target tracking algorithm based on improved Kernelized Correlation Filters

Visual target tracking is one of fundamental research of computer vision field and play an important role in the surveillance application, but it is also one of the difficulties due to the instability of the tracking scene. In this paper, we analyze the major drawbacks of the original Kernelized Correlation Filter (KCF) tracker which causes tracking failure when target experience complicated scenarios such as deformation, heavy occlusion and scale variations. In order to alleviate these drawbacks, we propose an improved KCF tracker, The tracker adopts a cascade classifier which composed by Multi-scale correlation filter and NN classifier. In each frame the tracking results are estimated by the relative variation of the target size. Experimental results of benchmark sequences show that the proposed algorithm has favorably performance against state-of-the-art methods of accuracy and robustness.