A Low Rank Regularization Method for Motion Adaptive Video Stabilization

Hand-held video cameras usually suffer from undesirable video jitters due to unstable camera motion. Although path optimization methods have been successfully employed to produce stabilized videos, the methods generally result in unintended large void areas in fast motion video sequences. To overcome this limitation, in this paper, we present a novel video stabilization algorithm which is derived from an optimization model consisting of a motion data fidelity term and two regularization terms: motion adaptive smoothness term and low rank term. Particularly, we design a motion adaptive kernel to measure neighbor motion similarity by exploiting local derivative information of dominant motion parameter, which is incorporated into the local weighted smoothness term to guide a motion aware regularization. Besides, the low rank property of neighbor motions is utilized to further improve the performance of stabilization. Experimental results show that the proposed method noticeably stabilizes a video, and it suppresses void areas effectively in fast motion frames.