Using Feature Point Detection And Matching Algorithm To Carry Out Multi-view 3D Reconstruction Research

Multi-view 3D reconstruction has been concerned for a long time,and has been applied in cultural relic digitization,medical and health care,industrial parts design,3D games and other aspects.The main methods of sparse point cloud reconstruction are incremental Structure-from-Motion(SfM)and global SfM.Because the bundle adjustment optimization of incremental SfM is easy to accumulate errors,resulting in scene drift,and has a certain dependence on the quality of the initial image selection,multiple bundle adjustments also increase the time of sparse point cloud reconstruction.Although the global SfM optimizes the relative translation vector and global position by building three views,and then optimizes the bundle adjustment,eliminating the error accumulation and scene drift,which makes the time of sparse point cloud reconstruction faster,the accuracy and integrity of sparse point cloud reconstruction is worse than the incremental SfM,as well as different image data sets will also lead to the failure of global SfM reconstruction.In view of the advantages and disadvantages of these two reconstruction methods,a hybrid sparse point cloud reconstruction method is proposed in this paper.First,the relative and global external parameter matrices are obtained by incremental calculation,and let them as initial values pass into the global objective function for optimization,then the accurate sparse point cloud is output by triangulation and three times bundle adjustments.The experimental results show that the hybrid SfM sparse point cloud reconstruction method proposed in this paper has more stable and robust reconstruction quality,higher point cloud accuracy and smaller error.The contents of this paper are as follows:1.The finite camera model,homography matrix,essence matrix and their solution are studied.This paper explores SfM point cloud reconstruction,introduces two view triangulation method and depth information estimation,camera pose solution and optimization method,such as Perspective-n-Point(PnP),direct linear transform,Bundle Adjustment,etc.,and the method of solving Bundle Adjustment(BA)is derived by an example.2.The similarity search is studied.The Scale-Invariant Feature Transform(SIFT)algorithm and the Fast Library for Approximate Nearest Neighbor(FLANN)rough matching are introduced in detail,and the difference between A Contrario-RANdom SAmple Consensus(AC-RANSAC)and RANSAC in the fine matching is mainly discussed.3.The method of SfM sparse point cloud reconstruction is studied.In the first chapter,the process and steps of incremental SfM and global SfM reconstruction are briefly introduced.Aiming at the advantages and disadvantages of the two reconstruction methods,a new hybrid SfM sparse point cloud reconstruction method is proposed.In the fourth chapter,the hybrid SFM sparse point cloud reconstruction method,reconstruction process and steps,incremental calculation of rotation and translation matrix,and the optimization of global objective function are described in details.In this paper,open source framework and open source image data set are used for hybrid SfM reseach,SIFT operator is used for feature detection,FLANN method is used for rough matching,k-dimensional tree(kd)algorithm is used to build feature descriptior search tree,and then using K-Nearest Neighbor(KNN)neighborhood to search for similarities,using AC-RANSAC to adaptively obtain robust correlation estimation matrix in fine matching and BA optimization.At the end of this paper,three reconstruction methods are compared and analyzed,the experimental results show that the hybrid sparse point cloud reconstruction method has some advantages in stability,robustness,accuracy and feasibility.