Object Tracking Based on Kernelized Correlation Filter with HOG and Illumination Invariant Features

Abstract

In order to improve accuracy and robustness of object tracking and meet the demand of real-time tracking, this paper presents a new tracking algorithm based on kernelized correlation filter with Histogram of Oriented Gradient(HOG) and illumination invariant features. At first, we calculated locality sensitive histogram of input image and extracted the illumination invariant features. Then we calculated HOG features, put illumination invariant features into the kernel circulant matrix based on HOG features. The tracking position is obtained by the responding confidence image, which can be quickly computed in the Fourier domain. Tests of many video sequences prove that the new algorithm has a better tracking performance than the traditional kernel circulant algorithm. The average tracking error of the new algorithm is 53 pixels lower than the kernel circulant algorithm, and the tracking precision is increased by 39%. As a result, the new algorithm can adapt to the conditions of illumination changes, pose variation and object occlusion.