Research On Multi-view Video Stitching Of Outdoor Scenes

Multi-view video stitching is a process that stitching multiple image sequences taken from different views into a wide view video.Multi-view video stitching widely used in surveillance systems and unmanned driving systems,so the research has practical significance.Multi-view video stitching often causes geometric distortion,and inconsistency of depth and motion(at scale and orientation).Focus on multi-view video stitching,the main work of this thesis includes four parts:1.Focus on geometric distortion,depth distortion and motion distortion caused by different views in an expansive camera system and focused camera system,we improve global projective transform,APAP(as projective as possible)and SPHP(shape preserving half projective)methods to stitch multi-view videos by considering angle constraints,and combine APAP and SPHP methods with angle constraints to improve further.2.A multi-view video stitching method based on angle constraint is proposed.Videos are first separated into static regions and moving regions.These two regions in videos are then stitched respectively,using the existing methods and those with angle constraints.After comparing the results and analysis,the combination method of APAP and SPHP with angle constraints for static regions.For the moving regions,geometric distortion is not obvious because of the moving objects are mostly in the middle of the image in our video data,therefore the global projective transformation and APAP and the two methods with angle constraints are used to stitch moving regions respectively.After comparing the results and analysis,APAP method with angle constraints is selected to stitch,and using optical flow to borrow feature information from the preceding frames to register.Videos are finally integrated with the stitched static and moving regions.3.Based on public datasets of in an parallel camera system,the existing methods(such as global projective transformation,APAP and SPHP)and the combination method of APAP and SPHP are applied to stitch respectively.After comparing the results and analysis the experimental results show that the combination method of APAP and SPHP has both high registration accuracy and shape preserving advantages.4.Based on captured videos in expansive camera system and focused camera system,the existing methods(global projective transform,APAP and SPHP),the combination method of APAP and SPHP,and these methods with angle constraints are applied to stitch respectively.The experimental results show that our method can sufficiently improve geometric distortion,and the inconsistency of depth and motion in multi-view video stitching.