Research On Real-time Three-dimensional Reconstruction Technology Based On Binocular Vision

Real-time 3D reconstruction based on binocular vision is based on the image taken by binocular camera as data input,through the visual algorithm to restore the three-dimensional model of reverse engineering.Stereo vision system equipment is cheap,and the reconstruction is speed and high precision,which is widely used in unmanned aerial vehicle inspection,mobile robots and so on.This thesis mainly introduces the three-dimensional reconstruction: camera parameters calibration,image stereo matching,point cloud registration and so on,for the commonly used algorithms were studied,and according to the specific scene proposed several improvements in the algorithm.The main contents of this thesis are as follows:1,The imaging model of the camera was studied in depth,the influence of camera distortion on the captured image and the necessity of camera parameter calibration were pointed out.For the calibration of the camera,this thesis uses the classic Zhang checkerboard method,the method is simple and high precision.In this thesis,the Harris corner detection algorithm is improved based on the characteristics of run-length coding and checkerboard symmetry.The subsequent experiments show that the improved algorithm can effectively filter the non-chessboard Grid corner,which is a good way to make up for the lack of Harris algorithm in a specific situation,making the use of checkerboard grid camera calibration operation easier.2,Stereo matching algorithm is the key to three-dimensional reconstruction technology,which to do all aspects of in-depth study in this thesis.First,the stereo matching criterion can improve the accuracy of stereo matching,and can greatly reduce the search range of matching point.This thesis introduces the common matching constraint criterion and its application scenario.Matching primitives and their metrics are the key points in the design of stereo matching algorithm,and also the key point to determine the performance of a stereo matching algorithm.Then,we introduce the problem of occlusion and sub-pixel parallax estimation in the calculation of matching metric and its solution.Then,this thesis studies three common stereo matching algorithms of SGBM,GC and BM.The experimental results show that the SGBM algorithm can take two aspects of performance and computational complexity.In our study,we also use this matching algorithm to obtain the disparity.3,Point cloud registration in access to large scenes or objects in the point cloud is essential,the current point cloud registration method,such as quaternion method,4PCS algorithm mainly through mathematical optimization point cloud registration,Those algorithms have high computational complexity and often results in local optimal error matching results after multiple iterations.In this thesis,a coarse cloud registration scheme is proposed.The initial value of the rigid body transformation is obtained based on the auxiliary camera calibration method.This scheme has the advantages of high precision and time consuming.After obtaining the initial value of the rigid body transformation,the ICP algorithm is used to optimize the value of the rigid body transformation several times,and then the point cloud precision registration is realized.4,In order to verify the visual algorithm in this thesis,we designed a 3D reconstruction vision system based on binocular camera.The system uses the binocular camera to obtain images of reconstructed object as the information input to obtain the depth information,then reconstructing the color three-dimensional point cloud.The system has a simple UI for user to operate easily.