Research On 3D Point Cloud Reconstruction Based On Multi-View Stereo

Visual 3D reconstruction has always been a hot topic in the field of 3D modeling.This kind of research combines multi-disciplinary knowledge such as multi-view geometry,computer vision,photogrammetry,etc.,and has high practical value in the fields of autonomous driving,medical imaging,cultural relic protection,etc.Based on the research of the theory and algorithm of multi-view visual reconstruction technology,this thesis designs a three-dimensional point cloud reconstruction system of monocular and multi-view,which completes the fine reconstruction of the detail contour and texture features of the scene structure by collecting the multi-view images of the scene.Firstly,this thesis introduces the basic theory of the camera imaging model,the projection transformation of spatial coordinates,and the three-dimensional reconstruction of polar geometry.A variety of camera calibration methods are analyzed.Finally,Zhang Zhengyou checkerboard calibration method is used to calculate the camera parameters,and the image distortion correction experiment is carried out.Then for the sparse reconstruction,the number of image feature detection and the accuracy of feature matching determine the reconstruction effect.Therefore,a feature detection method combining SIFT algorithm and SURF algorithm is proposed to improve the number of feature points,and the accuracy of matching is improved by reducing the matching threshold and eliminating the wrong matching.Based on the above joint feature detection method,Structure-from-motion(SFM)was used to estimate the position and attitude relationship of the camera in each image.Combined with triangulation and nonlinear optimal solution theory,the 3D sparse point cloud of the target scene under the condition of minimizing reprojection error was solved.For dense reconstruction and the final part,this thesis mainly analyzes the rest based on another perspective of the reconstruction(PMVS),based on the sparse feature point rest to build another model,using photometric consistency rest for another growth and filtering process,get the target scene of dense 3Dpoint clouds,through the experiment on the image and the experiment of 3Dreconstruction with open source image respectively,Good reconstruction results were obtained.In this thesis,the truth data of building scene obtained by 3D scanner is introduced to quantitatively analyze the overall reconstruction accuracy of this algorithm,which meets the general needs of 3D modeling.