Research On Image Registration Algorithm In Panoramic Stitching

Panoramic stitching is a cross-processing technology that integrates computer vision,computer graphics,and image processing.It is committed to obtaining a full-view image without introducing artificial traces.With its excellent viewing experience,panoramic images have been widely used in many fields such as aerospace,deep-sea exploration,security monitoring,target tracking,and live photography.Accurate image registration is an important prerequisite for ensuring the quality of panorama stitching.However,the current image registration algorithms for panorama stitching have many limitations.In addition,as fisheye lenses are widely used in panoramic stitching,severe lens distortion has further aggravated the difficulty of image registration.This thesis studies the image registration algorithm in panoramic stitching,and launches the thesis from two aspects of theory and application.First,the Random Sample Consensus(RANSANC)algorithm based on the rotation matrix model is proposed to solve the problem of filering the matching results of fisheye image feature points in the small baseline scene,and sufficient evenly distributed matching information was obtained for subsequent registration optimization.Then,through the research and analysis of the panoramic stitching projection method,this thesis proposes a local registration algorithm based on multi-rotation motion to complete the alignment of the partial projection of the image in the panoramic stitching.The main innovation and work of this thesis are as follows:(1)A RANSAC algorithm based on rotation matrix model is proposed.This thesis analyzes the principle of the traditional RANSAC algorithm based on the homography matrix model and the fundamental matrix model.Aiming at the problem of insufficient performance of the traditional RANSAC algorithm on the fisheye image,the RANSAC algorithm based on the rotation matrix model is proposed.It can effectively achieve the filtering of the matching results of the feature points of the fisheye image in the small baseline scene,and roughly estimate the rotation matrix between the cameras.When the camera intrinsic parameters have been calibrated,the rotation matrix is estimated by restoring the imaging light from the feature points to eliminate the influence of lens distortion,and the matching of the inlier points is filted by the angle of the light.(2)A local registration algorithm based on multi-rotation motions is proposed.Through the research of the single-view projection model in panoramic stitching,the main reasons for ghosts in panoramic stitching are analyzed.Aiming at the problem of insufficient image registration ability with a single projection parameter,a local registration algorithm based on multi-rotation motions is proposed.By dividing the projection space according to solid angles and individually adjusting the projection parameters of different areas,the same image camera can perform multiple rotations to achieve the local alignment effect for different levels of scenes.(3)Combining the RANSAC algorithm based on the rotation matrix model and the local registration algorithm based on multi-rotation motions proposed in this thesis,a panoramic stitching scheme directly applicable to fisheye images is proposed.The subsequent processing flow is unified by using camera parameters to perform standardized mapping of the fisheye image.Then through the RANSAC based on the rotation matrix model to complete the filtering of the matching results of the fisheye image feature points to obtain sufficient matching information.And through the local registration algorithm based on multi-rotation motion,the optimization and adjustment of the projection parameters of the fisheye image are directly completed;finally,the final panoramic image is obtained through image fusion.(4)In this thesis,a multi-level urban architectural scene is constructed through Blender 3D modeling software,and different camera baselines and different types of fisheye images are taken for experiment and analysis by controlling conditions.It is verified that the RANSAC algorithm based on the rotation matrix model can effectively achieve the filtering of feature points matching in fisheye images in the small baseline scene;the local registration algorithm based on multi-rotation motion can improve the local registration results of the image effectively;And the fisheye panoramic stitching scheme that combines the two algorithms can effectively complete the stitching of various fisheye images.