A registration-based tracking algorithm based on noise separation

Abstract Object tracking by image registration is one of important issues in image processing and computer vision. The traditional registration-based tracking algorithms are based on the optical flow constant assumption, and therefore are very sensitive to some challenging factors (such as illumination variation, pose change, partial occlusion and so on). In order to handle these challenges, this paper proposes a robust registration-based tracking algorithm. This algorithm models the tracked object by using a principle component analysis (PCA) subspace, adopts the idea of noise separation to handle the abnormal noise that appears in the tracking process, and designs the corresponding objective function. This objective function can be effectively solved by an iteration algorithm, and then the representation coefficients and motion parameters are obtained at the same time. The experimental results on some challenging video dataset demonstrate that the proposed tracking algorithm could achieve promising performance.