Research On Three-Dimensional Reconstruction Based On Binocular Stereo Vision

With the rapid development of computer science, it has been a hot research direction that machine vision replaced the human eye to complete the judgment and measurement of objects and realize the three-dimensional reconstruction of the information. Based on binocular vision,the method of three dimensional reconstruction uses two cameras to observe the object from different angles and utilizes the results of stereo matching to calculate the deviation of the image feature points. Finally, the object depth information is obtained to realize the three-dimensional reconstruction according to the principle of triangulation measurement. The basic process mainly includes image acquisition, camera calibration, image preprocessing,stereo matching, and three-dimensional reconstruction.Firstly, this paper completes construction for hardware platform of binocular vision and introduces the mathematical model of binocular vision. On the side of image acquisition, the multi-group image acquisition is completed by using the built vision platform and taking full account of the light intensity and the characteristics of the object. On the side of camera calibration, it is prepared for subsequent processing based on the principle of nonlinear camera imaging model, and camera intrinsic parameters, external parameters and the distortion coefficients are calculated by Zhang Zhengyou’s plane template calibration method. On the side of image smoothing, this paper introduces the Gaussian filtering and median filtering method, and makes comparative experiments between the gauss filtering and median filtering.On the side of edge detection, the Roberts operator, Prewitt operator, Sobel operator, Log operator, and Canny operator are analyzed, and further compared through the experiments. In the aspect of extraction for feature point, the traditional SURF algorithm is improved in this paper. Some feature points with less information content and relatively close distance are removed mainly using the information and distance constraints through the characteristic points detected by the traditional SURF algorithm so that the number of the feature points extracted is moderate and the aggregation of feature points is eliminated. In the aspect of feature points matching, the matching is accomplished using nearest neighbor matching algorithm based on the feature points extracted by improved SURF algorithm. Though comparing with traditional SURF algorithm about three different objects in two differentconditions, the results show that the improved SURF algorithm is robust to changes in brightness, and matching accuracy and the required time is better than traditional SURF algorithm. What’s more, at the time of the detection for a large number of feature points, some feature points with less information content and relatively close distance are effectively removed. In the aspects of three-dimensional reconstruction, three-dimensional coordinates of the object are calculated by the principle of three-dimension according to the principle of three-dimension, and finally the depth map in three-dimensional space is restored.