Study On The Methods Of Video Stabilization Based On Sparse Optical Flow

Videos are intuitive and informative, but for real video capture devices, due to the interference of external environment, such as hand-shaking, rough road, strong wind and other factors, the boring jitter often occurs in the capture videos. It is inconvenience for people to observe and analyze the information of videos. The digital video stabilization is to improve the video quality by removing jitter movement, which has been widely used in the various fields.In this paper, the motion relationship of neighboring images is calculated based on sparse optical flow,. then use different motion models and compensation models to implement digital video stabilization. The main work includes the following aspects:(1) The basic framework and classification of digital video stabilization are summarized, the advantages and disadvantages of different motion models are briefly introduces. Furthermore, the methods about motion estimation and compensation for 2D technology, the efficient image reconstruction algorithm for 3D technology and the local subspace constraints for 2.5D technology are described in detail, which are the advanced algorithms of digital video stabilization.(2) A real-time video stabilization method is proposed based on multi-resolution matching and improved Kalman filter. Firstly, the blur values for every input frame are computed, to determine whether the image is blurred by the threshold value. The traditional method of tracking sparse optical flow is used for non-blurred image matching, on the contrary, the multi-resolution sparse optical flow tracking algorithm used for blurred image matching. Then, the adaptive RANSAC algorithm is adopted to effectively eliminate the interference of the feature points on dynamic objects for motion model estimation, at the same time, the parameters of motion model are calculated by least-squares method. Furthermore, by improved Kalman filter algorithm to smooth the motion parameters, which can avoid the motion-lag phenomenon. Finally, the motion-compensating motion parameters are used for image reconstruction. For the videos with fast scanning and motion blur, the proposed algorithm can achieve good stabilizing results, and can finish video stabilization in real-time.(3) Another video stabilization is proposed based on local complete trajectories and grid warping. Firstly, the feature point trajectories through the entire video are obtained by sparse optical flow, at the same time, the forward-backward matching error is used to eliminate the wrong optical flows, and a coarse to fine method for dynamic features exclusion is adopted. Then, the local complete trajectories algorithm, which contains the construction of local complete trajectories matrix, the augment of incomplete trajectories and the smooth of local complete trajectories matrix, is picked to obtain the smooth vectors of feature points in every frame. Finally, the grid warping is used to reconstruct image scene. The method belongs to 2.5D technology, possessing the advantages of both 2D and 3D algorithms. It is possible to achieve desired stable videos.For the videos of different scenarios, our methods are compared with other video stabilization algorithms. The experimental results have shown that the proposed algorithms have high robustness and better stabilizing effect.