Research On 3D Reconstruction Technology Based On Stereoscopic Vision

In the field of computer vision, it has become a hot spot to reconstruct 3D scene model from 2D images. At present, 3D reconstruction can be carried out in three ways: geometric modeling, 3D scanner 3D modeling and 3D reconstruction based on image. In which, 3D reconstruction based on image has the advantages of high speed, convenience and low cost.The main content of this paper is from the two aspects of three-dimensional reconstruction of computer vision. The first is the 3D reconstruction of the scene based on binocular camera .The second is the reconstruction of the target based on the sequence image.Three dimensional reconstruction based on binocular stereo vision can be divided into several important parts: camera calibration, feature extraction, stereo matching and 3D reconstruction. The main steps are as follows: (1) The calibration of the two CCD cameras is completed by using Zhang Zhengyou checkerboard calibration method, by which we can obtain cameras' internal and external parameters. (2)According to the calibration results, we solve the re-projection matrix. (3) Extracting the SIFT features. (4) Stereo matching using epipolar constraint and KNN. (5) Using the matching relation to obtain the ocular disparity . (6)Obtaining the 3D point cloud of the object or scene by using the ocular disparity, the matching point coordinates and the re-projection matrix.On the basis of the SfM (Structure from Motion) 3D reconstruction technology, a new algorithm of 3D reconstruction is proposed. It is mainly composed of the following steps: (1) In the aspect of feature selection, the SURF feature is used instead of the traditional SIFT feature for feature extraction and matching in order to speed up the matching speed. The non - target region of the scene is masked by two - pass mask mechanism to eliminate the false matching between the noise feature points and features. (2) The essential matrix which can reflect the relationship between cameras is obtained by using the point to point matching of two cameras.Decomposition of the essential matrix to get the relative transformation between two cameras R and T, set the first camera coordinates for the world coordinate system. (3) By using the transformation matrix between the two cameras and the coordinates of each pair of matching points, the spatial coordinates of the target or the field are obtained by the method of triangulation reconstruction, and the initial point cloud is obtained. (4) After the initial point cloud is obtained, the residual image is reconstructed by the incremental method, and the new point of view is fused with the previous 3D point cloud to get the point cloud of the object in the multi view image. With the increasing of the image, the error will accumulate gradually, and there is the phenomenon of error accumulation in the incremental 3D reconstruction. In this paper, the central camera in the camera topology is used as the world coordinate system to eliminate the error caused by the excessive error.The experimental results show that the average error of the calculated values and the measured values of the 3D reconstruction based on binocular stereo vision is less than 1.72%.The three-dimensional reconstruction based on the sequence image use SURF feature extraction and matching.It's speed has been improved. The mask is used to process the feature extraction to reduce the mismatch phenomenon. According to the camera topology, the world coordinate system is set up to realize 3D reconstruction. The experimental results are more intuitive and more realistic. Sequence image through the algorithm to reconstruct the three-dimensional coordinates, the calculated value and the actual measured value in the accuracy of the average error of not more than 0.93%.