Study Methods Of Video Stabilization Based On Adaptive Local Subspace Of Feature Points Classification

In the actual shooting environment, the unstable videos result from vehicle vibration or other factors of external environment. Shaking videos not only result in bad views but also introduce negative effects on visual processing work. Digital video stabilization technology uses image processing algorithms to remove the jitters from shaking videos, and has advantages of better accuracy, cheap cost over the mechanical and optical video stabilization technology. Therefore, the Digital video stabilization technology has the fatal research significance.This paper focus on the basic theory and main algorithms of digital video stabilization. The detailed research content is mainly composed of the following aspects:(1) The technologies of digital video stabilization are summarized briefly. Firstly, the basic principles and methods of the 2D technology are presented. Secondly, the algorithms of structure from motion and content preserving warp in the 3D technology are analyzed and expounded. Thirdly, the methods of outlier detection and optimal camera paths in the 2.5D technology are introduced. Finally, this paper gives the comparative advantages and disadvantages of various technologies.(2) The quick camera motion will lead to trajectories dropping quickly, and the moving foreground will result in inaccurate estimation of the motion vector. The paper proposes a video stabilization method based on adaptive local subspace of feature points classification to solve the problems above. Firstly, the feature point classification method is based on global motion estimation to exclude outliers efficiently. Secondly, multiple low-rank local subspaces constrain the constructed trajectories matrix with an adaptive size. In that way, the method preserves the geometry relationship of local feature points. Thirdly, the homography consistency alleviates the abrupt inter-segment transition to make frames joint smooth. Our method has better performance than the other state-of-art methods and can effectively stabilize shaking videos with quick camera motion, moving foregrounds and scenes depth variation.(3) This paper gives the method about stereoscopic video stabilization based on joint local subspace of feature point classification. Firstly, this paper proves that low-rank local subspace also holds for stereoscopic videos. The feature trajectories from left and right share the same local subspace. Secondly, the joint local subspace maintaining the consistency between left and right is used to constrain trajectory matrixes. Thirdly, the method of synthesis stereoscopic videos verifies the effect of video stabilization. Experiments show that, our method can achieve better video stabilization results under a realistic stereoscopic effect.Simulate experiments with a great quantity of different scenes videos are conducted to evaluate our methods and compared with other state-of-art methods. Experiments demonstrate that our methods approach better performance.