A Video Stabilization Algorithm Based On Reprojection Using Trifocal Tensor

With the rapid development of information technology,camcorder devices have been widely popular,which makes the videos ubiquitous in people's life.Videos have become an important way for people to get information resources thanks to their rich content,visual expression and easy access,storage and transmission.However,due to the jitter of a image carrier,the collected video jitters.Consequently,The jitter of videos introduces unpleasant visual effects and negatively affects on the follow-up use of the videos,such as object tracking,and video compression.In order to obtain more comfortable visual experience and facilitate the subsequent use of video,we propose a feature trajectory video stabilization algorithm based on trifocal tensor to have solved the three problems as follows:(1)Aiming at the problem that the existing feature trajectory video stabilization algorithms can not take into account the trajectory length,robustness and trajectory utilization simultaneously,a feature trajectory stabilization algorithm using trifocal tensor is proposed.This algorithm uses trifocal tensor transfer to construct the long virtual trajectories.Then it defines the stabilized views by smoothing the virtual long trajectories.After that,it smooths the original feature trajectories by reprojecting the original feature points to the stabilized views.The reprojection is implemented using trifocal tensor.At last,the stabilized frames are rendered by wrapping the original frames using Content-Preserving Wraps.(2)On the issue that the video stabilization algorithms have difficult in doing with strong moving object occlusion at present,we propose the algorithm based on K-means to remove feature points of moving objects.This algorithm uses speed and position of feature points as the input for K-means.It clusters the feature points on a frame into K clusters,then fuses the clusters of background feature points and the clusters of feature points of moving object respectively by epipolar geometry constraint.Afterwards the algorithm gets the background cluster based on the compactness of feature points.The algorithm uses K-means clustering on all frames,then takes advantage of the proposed criterion about the cluster results of all frames to get rid of feature point of moving objects.This method can avoid smoothing the feature point trajectories of moving objects and the severe errors of trifocal tensor computation which is generated by feature point trajectories of moving objects.Therefore it can avoid the distortion of stabilized frame.(3)In order to verify the performance of the proposed algorithm,we design contrast experiments.Through subjective feeling of three people for stabilization results of two kinds of algorithms,we findout the better result.We divide the processed videos into several categories and use the statistics of evaluation results for every category to reflect the algorithm stability in the situations lacking long trajectories or containing rolling shutter distortion,etc.Experimental results show that,The proposed algorithm has the low requirement for trajectory length,high trajectory utilization rate and good robustness.It not only can deal with the non-planar structure scenarios,lack of parallax and rolling shutter distortion,but also can handle with the scenarios lacking of long trajectories,which is caused by quick panning of camera,excessive jitters,motion blur.And the algorithm is also suitable for dealing with strong occlusion caused by moving objects.