Research On Stereo Matching Technology For 3D Reconstruction Of High-resolution Remote Sensing Images

With the rapid development of national remote sensing satellite earth observation technology,obtaining a large number of high-resolution remote sensing satellite stereo images is no longer a problem.The 3D reconstruction of remote sensing images is playing an increasingly important role in the fields of urban planning and disaster warning.Among them,the stereo matching technology of remote sensing images is one of the key technologies of 3D reconstruction.This technology obtains image depth information by calculating the horizontal deviation of corresponding points in images taken from different angles.It is widely used in fields,such as autonomous driving,three-dimensional reconstruction,urban planning,and disaster warning.Since the stereo matching algorithm is sensitive to noise,shadow areas,discontinuous disparity areas,and weak texture areas,in the process of matching remote sensing images,mismatches will occur when the matching costs are aggregated,and there will be many in the generated disparity map.Streaks and noise cannot be directly applied to the subsequent 3D reconstruction.Therefore,this thesis improves and optimizes the problems of the semi-global matching algorithm(sim-global matching,SGM)in the stereo matching algorithm.The main research contents are as follows:(1)Research on the remote sensing image block strategy based on quad-tree structure.Because the remote sensing image has a large width when shooting,the length or width of an image is often hundreds of kilometers.It will be very time-consuming to perform feature detection and preprocessing on the remote sensing image.In order to solve this problem,a method is proposed.Remote sensing image block algorithm based on Morton coding quad-tree structure.First,Morton code is used to encode the remote sensing image;then,a quad-tree structure is established according to the encoded remote sensing image,and then the remote sensing image is divided into blocks;finally,the feature detection is performed by obtaining the divided image.Compared with the average block algorithm and the uneven block algorithm,this block algorithm can extract more feature points.(2)Research on the image generation algorithm of remote sensing epipolar stereo pair image based on geometric constraints.The stereo matching process of remote sensing images is to search for points of the same name in two-dimensional space,which will occupy a lot of aggregation space during the matching cost aggregation stage,and the calculation of matching cost is large.To solve this problem,a remote sensing image stereo based on geometric constraints is proposed.The matching epipolar image generation algorithm reduces the search range from two-dimensional to one-dimensional space during the aggregation process of matching cost of stereo matching.First,overlapping area analysis and regular grid division are used to cut up the remote sensing image;then,the improved SURF algorithm is used to perform feature detection and matching on each small image block to obtain the point of the same name of the stereo pair;finally,the obtained the point pairs of the same name are used to calculate the projection basic matrix and the affine basic matrix of the remote sensing image,and then use the improved eight-point method that adds the residual of the matrix and the second derivative of the rotation angle to determine the precise matrix to make the epipolar image,and then use the epipolar line The image undergoes subsequent semi-global stereo matching to obtain a disparity map.(3)Research on semi-global stereo matching algorithm combined with accelerated robust features.The remote sensing image SGM algorithm has problems that it is sensitive to noise and does not consider the correctness of the aggregation path direction when the disparity discontinuity and weak texture regions are aggregated for matching cost.This results in stripes in the final generated disparity map,resulting in low matching rate.In order to solve these problems,a semi-global stereo matching algorithm combining speed up robust features is proposed.First,the speed up robust features algorithm is used to get the feature matching points and their main directions,and then use the fast nearest neighbor search algorithm to eliminate the wrong matching points;then,because the Census transform has robustness and good matching effects to the lighting changes,the Census transform is used to calculate two images For the matching cost of remote sensing images,the main direction of the feature points is used to adjust the path weight of the SGM algorithm in different aggregation path directions;finally,the initial disparity map is refined to remove the noise and stripes in the disparity map,and the final disparity map is obtained.