Research On 3D Reconstruction Based On Binocular Vision

The 3D reconstruction is one of the main contents in machine vision researching field, studying the methods how to restore the depth information of scenery from two or more two-dimensional images. Binocular stereo vision technology based on two images is the hot-spot research imitating the working of eyes. It makes use of geometry principles to analyze and compute the parallax of image pairs, then obtains the depth information of scenery. This thesis studied the methods and key techniques of 3D reconstruction based on binocular stereo vision. The main contents are as follows:1) A hardware platform for stereo images acquisition was built, which consists of PC, capture card,1394 bus, binocular camera and bracket. The software framework for camera control and image capture was developed on the basis of SDK.2) The method of plane calibration template invented by Zhang Zhengyou was adopted to calibrate the camera to get camera parameters and distortion factors. Those captured images were processed by methods of filtering for noise reduction, sharpening, and histogram equalization in the preprocessing.3) We adopted the matching method based on feature points and made a contrastive analysis of feature detectors commonly used. And then, we selected SAD as the evaluation of correspondence and took advantage of the constraint conditions including epipolar line, uniqueness and disparity range etc. to reduce the search range and mismatch.4) Under the premise of getting intrinsic parameters and extrinsic parameters of camera and the right matchers, we computed the depth of key points according to the principle of stereoscopic vision measure, and then got the corresponding information in 3D space. We used OpenGL as visualization tools to display the reconstructed three-dimensional data.