Research On Image-based 3D Reconstrucion And Measurement Technology

Image-based 3D reconstruction and measurement is a fundamental and important problem in computer vision.It is the basis for computer analysis,understanding,and manipulation of 3D targets.However,due to matching ambiguity and error magnification in triangulation,there are plenty of noises and errors in image-based 3D reconstruction,which seriously restricts its practical application.Therefore,providing accurate 3D reconstruction algorithms has high theoretical significance and practical values.This dissertation is dedicated to improving the accuracy of image-based reconstruction and measurement,which has carried out specific theoretical and experimental research from three parts: monocular vision,binocular stereo vision,and point set registration.The main works are summarized as follows:Firstly,a pose-free GVC algorithm based on image center registration is proposed.For images with large angle changes and weak textures,GVC can reconstruct effectively.However,it has strict requirements on the optical axis direction.Because the optical axis of the camera is invisible,the optical axis direction cannot be accurately controlled.In order to remove the constraint of GVC on the optical axis direction,this paper proposes a view-free GVC algorithm,which aligns the optical axis direction by aligning the image target center.In addition,the proposed algorithm uses robust line features as the basic features of image target center positioning and converts them into the unique body diagonal intersection.Among them,the center of the image target and the center of the space target correspond to each other through the body diagonal,which ensures the accuracy of the algorithm.Through the elliptical orbit reconstruction experiment with the short axis of 100 meters and the long axis of 200 meters,the proposed method is more effective and accurate than the existing methods in terms of the lack of camera pose,large angle of view change and weak texture reconstruction.Secondly,a binocular stereo vision reconstruction method based on lidar correction is proposed.There are many noises in binocular stereo vision and the depth measurement declines with the square of the distance.And the decline of measurement accuracy is an inherent characteristic of binocular vision,which cannot be solved by algorithms.In this dissertation,lidar with accurate distance measurement is used to correct the depth of binocular vision,thereby the depth accuracy of binocular vision can be guaranteed by lidar.For the registration of binocular and lidar,maximum likelihood estimation and EM algorithm are adopted to estimate the transformation between two point sets.For the fusion module,two reconstruction methods are proposed: the nearest neighbor interpolation and the hidden Markov interpolation.Experiments on 3D lidar and 4-layer lidar verifies the effectiveness of the proposed algorithm in suppressing reconstruction noise and improving the accuracy of binocular reconstruction.Thirdly,a non-rigid point set registration method based on high-dimensional representation is proposed.Point set registration is the most important step in the reconstruction scene stitching.Existing registration methods match two point sets according to the global and local features.Due to distortion,noise,rotation and the irregularity of the point set,there is ambiguity in global and local features,which affects the registration effect.In this paper,the relative distance is added to point coordinate,and the stability of the relative distance in the distortion guarantees the robustness of the proposed algorithm.Meanwhile,the high-dimensional coordinate is used to calculate the membership probability,and the accuracy of the membership probability is ensured by the magnification of the distance of non-corresponding points with highdimensional coordinates.Experiments on 2D simulation point sets,3D simulation point sets and the real 3D data demonstrate that the proposed algorithm achieve results superior to those of state-of-the-art methods in terms of deformation,noise and rotation distortion.