Research On Image Homography Estimation And Inter-Frame Stabilization Algorithm In Video Stitching

Video stitching technology combines multiple overlapping videos shot at the same time to form a video with a broader vision,which is widely used in virtual reality,security monitoring,medical treatment and other fields.To meet the demand for real-time performance in the application scenarios of video stitching,existing image stitching algorithms typically apply fast matching algorithms such as ORB.However,the accuracy of fast matching algorithms is relatively low,which results in less accurate estimation of the homography transformation matrix and affects the quality of the stitched image.Moreover,if the stitching image is directly encoded to form a video,it will cause obvious jitter in the resulting video,which is not conducive to human observation.This dissertation focuses on the problems of homography transformation matrix estimation and video stabilization mentioned above.To address of low matching accuracy in fast matching algorithms leading to inaccurate estimation of the homography transformation matrix,this dissertation proposes an improved fast matching algorithm based on BEBLID and a homography matrix correction algorithm based on overlapping regions.The improved fast matching algorithm based on BEBLID first selects seed points from the feature points extracted by the BEBLID algorithm.Then,adaptive local affine transformation verification is performed in the neighborhood of the seed points,and matching results are iteratively filtered.This algorithm effectively improves the matching accuracy while ensuring real-time performance.In the homography matrix correction algorithm based on overlapping regions,after the first image transformation,the homography transformation matrix is re-estimated in the overlapping area to correct the homography transformation matrix calculated after the initial matching,which enables the image to be more accurately transformed onto the plane of the reference image.Since there are few existing image quality evaluation metrics that evaluate the alignment degree of spliced images,this dissertation also proposes a new metric called image frame projection error to evaluate the alignment degree of pixel points after image projection transformation during the stitching process,which is used together with PSNR and SSIM to evaluate the quality of stitching images.To address the jitter problem in stitching videos,this dissertation proposes a video stabilization algorithm based on PWC-Net optical flow grid tracking.The PWC-Net optical flow between any two adjacent frames is obtained and divided into a grid,and the motion of grid points is tracked for global motion estimation.Then,filtering and smoothing is applied to remove jitter components in motion,and a compensation motion model is obtained according to the smoothed motion trajectory,and the corresponding transformation is compensated between any two frames of the video,thus achieving video stabilization.The experiments show that compared to the stabilization algorithm based on LK optical flow,the algorithm proposed in this dissertation has a significant improvement in video stability evaluation,with the highest increase of 5.92%.