An improved object tracking based on spatial context

This paper proposes an improved object tracking method based on spatial context of image to improve the accuracy and real-time of object tracking. First, the image is randomly sampled around target at current frame. We compare each sample with template image using the kernel method in Fourier domain so that we will obtain the location of the maximum response. Then, in this position, the pixel similarity is summed by the weighted Gauss function within 10*10 sub-window, and the location of the maximum similarity in all sampling regards as the best tracking position. The experimental results demonstrate that the tracking speed is obviously improved because Fast Fourier Transform(FFT) is adopted in algorithm. Tracking algorithm runs at about 100 frames per second on i5 machine. Tracker precision reaches about 90% at a threshold of 50. Extensive experimental results show that the proposed algorithm outperforms favorably against state-of-art tracking methods based on kernel method in many complex conditions.