Research On Image Feature Point Matching Method In 3D Reconstruction

Three-dimensional reconstruction is the process of automatically calculating and building a 3D geometric model of the image data of the real scene collected by the camera,and trying to recover the information of the 3D space from the 2D image information.In recent years,3D reconstruction has been paid close attention by researchers in the fields of virtual reality,intelligent driving,smart home,medical treatment,cultural relics protection,etc.Therefore,the research on the key technologies and difficulties in the process of 3D reconstruction has important practical significance.Image feature point matching is a crucial step in the process of 3D reconstruction,and the accuracy of the matching results directly affect the subsequent 3D reconstruction effects.On the basis of sorting out the research results and data of researchers in related fields,this paper focuses on the key problem of image feature point matching in 3D reconstruction,and researches on improving the performance of image feature point extraction and initial matching in 3D reconstruction,and optimizing the method of removing mismatching points in 3D reconstruction.The main contents of this paper include:(1)Aiming at the problem that the descriptor dimension of the traditional SIFT algorithm is too high and the reconstruction time is too long,this paper proposes an improved SIFT feature point extraction and initial matching method.On the basis of image preprocessing,this paper first uses the SIFT algorithm to detect the specific position of each key point,and at the same time Sobel operator is used to calculate the gradient of the image,giving each pixel a new gradient definition;then the original square grid is replaced by eight concentric ring structures,and a 64-dimensional descriptor is generated based on the concentric circle neighborhood.Finally,an initial matching method of feature points based on neighborhood voting is proposed in the initial matching strategy,and the matching points are searched by distance and direction constraints.Experiments show that the proposed method can reduce the initial matching time and obtain high initial matching accuracy,and it still has good adaptability to the image transformation of complex scenes.(2)Aiming at the problem that there are a lot of overlapped areas in the 3D scenes captured by the camera and that the traditional RANSAC algorithm's removal effect is not ideal due to the increase in the outlier ratio,a simplified RANSAC algorithm is proposed in this paper.All the last 20% initial matching points obtained by neighborhood voting are deleted as unqualified points,and deal with the remaining points,so as to increase the ratio of interior points.The experimental results show that the improved RANSAC algorithm has better adaptability to the situation where there are overlapping areas between the images in the remote sensing image and the actual shooting scene with high matching difficulty,and it improves the efficiency of eliminating the mismatching points and the accuracy of the transformation matrix.(3)In order to verify the performance of the image feature point matching method proposed in this paper in 3D reconstruction,the experiments of initial matching based on Improved SIFT algorithm,mismatching point elimination based on optimized RANSAC and 3D sparse reconstruction are carried out respectively.Finally,the 3D sparse reconstruction of the image is realized,and the sparse point cloud with more reasonable distribution and good reconstruction effect is obtained.The experimental results verify the feasibility and effectiveness of the image matching method proposed in this paper for the reconstruction results.This thesis has 41 pictures,9 tables and 82 references.