An Image Stabilization Algorithm Using Feature Trajectory Augmentation

Driven by the rapid development of science and technology,photography has become a part of people's lives.People use video to record and transmit information.Video is widely used as a carrier of information in fields such as life and industrial science and technology.However,sometimes due to the influence of external factors,there is a certain amount of jitter which often results in the video shot,which is not conducive to the transmission of information and follow-up treatment work.Therefore,in order to improve the quality of video,this paper proposes a video image stabilization algorithm based on feature point trajectory augmentation,studies and solves the following three problems:(1)In view of the existing feature point trajectory stabilization algorithms,which can not take the four aspects: trajectory length,trajectory utilization,robustness and time complexity into account,this paper proposes a video image stabilization algorithm based on feature point trajectory augmentation.Firstly,the algorithm improves the feature point selection strategy: take the feature point trajectory length as one of the parameters of information measurement,and use information metric evaluation function to filter the feature points;then the trajectory growth is achieved by using the iterative approximation principle of low-rank matrix,so as to satisfy the algorithm requirements of trajectory smoothing;finally the velocity vector between the coordinates of the smoothed feature point and the initial feature point is used as a guideline to guide the deformation of the mesh of the original frame to generate a stable frame.(2)On the issue of moving objects,the advantages of K-means clustering and wavelet clustering are integrated.The K-means-wavelet fusion clustering algorithm is proposed to remove the feature points of the foreground moving target.The algorithm selects the speed,coordinate and RGB value of the feature points as the observation data of each feature point to determine the feature point distribution type,then selects the K-means clustering algorithm for the spherical distribution feature points to eliminate the feature points,or selects the wavelet clustering algorithm for the non-spherical feature points to achieve the feature point removing.Application of wavelet clustering does not require similar clustering centers between clusters,so it can well deal with non-spherical clustering,which can improve the shortcomings in the K-means clustering,but because of the wavelet clustering algorithm basing on frame difference and grid,there is no algorithm for the effect of standard spherical clustering.Therefore,the clustering algorithm is combined with wavelet clustering algorithm to increase the processing range of video image stabilization.(3)By using comparing experiments,we verify the effectiveness of our algorithm and compare them with the three proposed algorithms which have good image stabilization performance: the algorithms of video image stabilization based on subspace constraint,the video image stabilization algorithm based on polar geometric re-projection,the feature trajectory video stabilization algorithm based on triple tensor.And the jitter video is divided into five categories according to the stability,spatial distortion,temporal distortion,time complexity,to conduct a better result.Experimental results show that the video stabilization algorithm proposed in this paper has strong robustness,high trajectory utilization and appropriate time complexity,which can well deal with scene changes with depth of field,fierce jitter,vagueness and the presence of one or more foreground motion targets in video.