Binocular Stereo Vision-based Three-dimensional Image Reconstruction

In recent years, cities have been conducted in the research and application of Digital City. aerial photographic surveying vehicle in which three-dimensional laser scanning as the primary means to access information, which is good to acquire surface data and the information on both sides of the road. But for the reconstruction of the indoor display objects, you need to use three-dimensional visual approach. Aimed at this practical application, in this dissertation, based on analyzing the existing research works of 3D vision measurement, a set of 3D image reconstruction based on stereo vision system is put forward, in order to find a high accuracy and fast solution.3D reconstruction using stereoscopic vision method, in computer vision is defined by two or a number of two-dimensional images to restore the visible 3D surface geometry of object. The three-dimensional reconstruction is a prerequisite for the establishment of a correspondence between the two images. In the establishment of such relations as an important constraint conditions is called epipolar constraint—the same space points in the projection of two images which is only satisfied by geometry relationship. Calibration of three- dimensional reconstruction is the key technologies. Firstly, linear and non-linear camera model are established; some methods of camera calibration. Experiment of camera calibration with non-linear camera model is accomplished.This paper presents a stereo vision method to obtain information of 3D objects. First, two cameras were used to obtain the images of the same object between the two views with few differences. Harris operator was used to extract the corner points as the feature. The matching gray was used to calculated the relevant points. RANSAC method was used to estimate the fundamental matrix, then restored epipolar geometry, The early matching sets and the epipolar geometry were used to search the exact points in one-dimension. And then the corresponding images of depth map was calculated, 3D points cloud information of the object was recovered. Results of real experimental data demonstrate this method.