3D Reconstruction Based On Binocular Vision Research And Implementation

3D reconstruction based on binocular vision is an important research in the field of computer vision. It has been widely used in precision measurement, robot navigation, virtual reality and other aspects. It has a very important practical significance and theoretical research value to build the 3D realistic modeling. In this paper, in order to improve the automation of binocular vision system, this paper has researched deeply on the image preprocessing, the camera calibration, feature extraction and matching,3D reconstruction and a series of key issues.Firstly, this paper analyzes the epipolar geometry and its solver methods, as well as the optimization to basic matrix. Secondly, we use the image from the different positions of calibration plate to calibrate the camera. The calibration result show that under the pixel level and the sub-pixel level have different effects on the last accuracy. Then comparing the results of the calibration by analyzing high-resolution images with low resolution, draw the camera itself distortion factors will adversely affect the accuracy of camera calibration, especially in the high-resolution images under this influence will be more prominent.Again, this paper has achieved good results in the match by using our method for any texture, checkerboard. For the non-coding point that can’t distinguish each other directly, this paper use the way of ellipse fitting to locate them after extracting the non-coding point. Due to the quantify of pixels, solver error of the basic matrix, rounding errors during the ellipse fitting and other factors, we find that the match point locate in the error band of the epipolar line corresponding. There must be a lot of one-to-many situation if only using two image. In order to avoid this situation, this paper use three images to find the matching points at the intersection of the error bands. The results show that not only can ensure the number of matching points, but also can ensure its uniqueness.In addition, this paper presents a way for 3D reconstruction by colored lines. By the way of block division to construct a new image of variance, quantify the each pixel, we can simply extract the line area at the same time suppressing noise. we express every line using the different function. This paper gave the straight line and B spline that depending on the curvature to choose which one. Then, we can obtain the sub-pixel precision intersection coordinates by the way of solving equations. Matching the intersection by the way of location and the surrounding color information. Evaluation of reconstruction precision by the flatness and the roundness. The results show that, by using this method, its reconstruction accuracy can reach millimeter.Finally, using the matching points to reconstruction. This paper obtain a good reconstruction results from the display of point cloud and surface.