Study On The Methods Of Digital Video Stabilization Based On Epipolar Geometry

With the rapid development of science and technology,more and more electronic products get the favor of people,such as digital camera,smart phones,video cameras,and so on.These electronic products become indispensable part of the people's live,work and entertainment.But when people shoot a video,video is be affected by the subjective or objective factors leading to video jitter.The shaking video will not only make people visual fatigue,but also will affect video follow-up studies.Therefore,digital video stabilization technology has drawn the attention of the more and more scholars.Methods based on epipolar geometry can reduce the interference of a moving object for camera motion estimation.In this paper,on the basis of reading a large number of literatures at home and abroad,the following several aspects of research work have been done:(1)This paper firstly introduces the basic framework and mathematical model of the digital video stabilization algorithm,and summarizes the video stabilization algorithm at home and abroad.The algorithm principle and advanced algorithm of 2D,2.5D and 3D algorithm are described,respectively.(2)In this paper,digital video stabilization algorithm based on the dense optical flow and epipolar geometry is proposed.The algorithm firstly estimates optical flow between two frames,then the matching point of each pixel is obtained.Secondly,all matching points are used to classify into two types based on the epipolar constraints,and the pixel of outlier type would be compensated based on homography.Finally,pixel trajectories could be smoothed by optimizing objective function.To ensure the image consistency of video,the algorithm proposed that the overlapping window is needed.(3)In this paper,digital video stabilization algorithm based on local subspace of FAST-ER feature points is proposed.The algorithm firstly extracts FAST-ER feature points on the first frame of video image.Secondly,Lucas-Kanade optical flow estimation algorithm based on pyramid model is used to match points for the subsequent image frames,then outlier points are judged based on epipolar constraints.Then,based on local subspace constraint,trajectory matrix is constructed.And smooth feature trajectories to get smooth feature trajectories.In the end,warp the original image sequences and output stable video streaming.For the videos with different scenarios,such as fixed scene,the scene with changing depth,camera zoom video,the scene with foreground movements and so on,this paper had done a lot of simulation experiments.In addition,the comparison of performances between the proposed methods and advanced algorithms was made.Experiment results showed that both of the proposed methods have good stabilization effect.