Research On Stereo Matching And Surface Reconstruction Algorithm Based On Binocular Vision

As one of the important branches of computer vision,binocular stereo vision has been widely used in virtual reality,games,archaeology,medical cosmetology and many other fields.It uses two cameras to acquire images of the same scene from different angles,obtains the coordinates of each pixel in three-dimensional space by calculating the parallax of the real object points in the two images,and finally reconstructs the surface of the points cloud to recover the real 3D scene information.Based on the study of the principle of binocular vision,this paper focuses on camera calibration technology,local stereo matching algorithm,surface reconstruction of scattered point clouds and multi-view registration.The main work of this paper is as follows:1.The three-dimensional reconstruction algorithm based on binocular vision is proposed,which includes four modules: image acquisition,camera calibration,stereo matching and surface reconstruction.2.Based on the study of the geometric principle of camera aperture imaging,the calibration experiment of binocular camera is carried out using planar chessboard lattice,and finally obtain the internal and external parameters of the camera,which lays a foundation for stereo matching and surface reconstruction.3.Based on the existing local stereo matching algorithm,the calculation model of matching cost is optimized,and the penalty factor in weighted guided filtering is improved by using Laplasse operator.The weighted guided filter with edge preservation property is applied to the process of matching cost aggregation,and the disparity map with higher precision is output.4.The surface reconstruction algorithm of scattered point clouds is designed and implemented.In order to obtain the realistic object structure,the spatial index structure based on kd-tree is constructed firstly.And then the Delaunay triangulation algorithm based on plane projection is proposed to optimized the search range of adjacent points.Finally,The holes in the surface model are filled with the wavefront method.5.An improved ICP point cloud multi-view registration algorithm based on curvature feature extraction is proposed.It can effectively reflect the intrinsicgeometric properties of objects and improve the speed of registration algorithm.It can make up for the missing side information of binocular camera when acquiring 3D point cloud.