Color-saliency-aware correlation filters with approximate affine transform for visual tracking

Aspects like deformation and occlusion are still the challenge cases which will result in failures of visual tracking. Many existing correlation filters (CFs) try to fuse the color information to improve the performance but ignore the sensitivity of color information to the background interference. For this case, we propose a color-saliency-aware correlation filter which exploits the color statistics as the model of image boundary connectivity cues. The proposed method limits the drift of correlation filter because of the saliency proposal. In addition, the bounding boxes of CFs absorb too much background information which can easily lead to the tracking failure. To solve this problem, we also present a decoupled-Fourier-Mellin (DFM) transform which is related to the independence of scale variations in log-polar coordinates. In addition to the rotation angle, the proposed DFM can also gain the scale factors of both horizontal and vertical directions, and the larger search space (5-DoF) is closer to the upper bound of object masks. Ultimately, multiple popular benchmarks demonstrate the superiority of our tracker. Compared with the current advanced CFs, our method achieves better performance, which is of great significance for continuous tasks requiring high DoF information, such as manipulator visual servo.