Stereo Matching Of Remote Sensing Image Based On Affine Invariant Feature

With the continuous improvement of satellite technology,the cost of remote sensing images acquisition is gradually reduced,and the resolution and other parameters of remote sensing image are gradually improved,how to obtain more useful information from remote sensing image has become an important research area.In many research fields,3D reconstruction has been paid more and more attention.There are many advantages to 3D reconstruction based on remote sensing images,such as large acquisition range and good real-time performance of remote sensing images,so there are more and more researches in this area.This thesis pays attention to the key step of stereo matching in 3D reconstruction of remote sensing images,and improves the algorithm in various aspects in order to achieve the goal of optimizing the stereo matching results of remote sensing images.Because the error of the disparity map results is proportional to the elevation error of the 3D reconstruction,this step is very critical.The research on stereo matching of remote sensing images is divided into the following parts.First,this thesis chooses the currently popular methods that are more suitable for stereo matching of remote sensing images.Based on previous studies,this algorithm is refined and optimized,and key steps are adapted for remote sensing images.After investigation and research,because of the requirement of matching accuracy,the local stereo matching of remote sensing images is not suitable,and the stereo matching method based on confidence is used.The research shows that this algorithm has unique advantages in stereo matching of urban remote sensing images.Therefore,based on the algorithm framework in this part,this thesis focuses on the pre-processing of remote sensing images and the selection of cost function.The final disparity results of remote sensing images are removed in occlusion areas and most plane are smoothed.Secondly,in order to solve the parallax expansion problem of the roof and the lack of side wall matching in the remote sensing image that cannot be solved by the traditional stereo matching method,the algorithm innovation is carried out in this thesis.These improvements finally make the parallax of the building roof area no longer have a large-scale expansionphenomenon,and at the same time,optimizes the plane void problem of the roof and the ground.It also enables the building side wall matching results to be improved from scratch.For the optimization of the roof,this thesis uses the steps of remote sensing image segmentation,segmentation block evaluation,distrust point removal,and plane fitting and filling to optimize.For the part of the side wall,this thesis carried out an experimental method of secondary matching.The main steps are extraction of the side wall of the building,matching of the affine change of the side wall,and backfilling of parallax results.Finally,this thesis designs innovative evaluation indexe for stereo matching of remote sensing images,and uses the method of sparse edge point matching of left and right images to verify the matching accuracy rate.This indicator partially solves the problem that stereo matching of remote sensing images has no accurate value.The index provides guidance to the improment.Finally,a large number of experiments in various aspects are carried out at the same time to verify the rationality and innovation of the algorithm.