Research On The Method Of 3D Reconstraction Based On Binocular Stereo Vision

As the same as human vision system, vision system is playing an important role in computer outside information collecting system. Vision technology is one of the hot topics in robot field recently. Just like humans, vision is one of the most important sources to know environment. This thesis has studied several key technologies of binocular stereo vision. The binocular stereo vision is the method that observes the same scenery between two different angle, in order to get the pictures on the different visual angle. By the image geometry theory, calculate pels's warp on two images, finally get the three dimensional information of the environment.The binocular stereo vision technique is achieved by capture of images, camera calibration, information image pretreatment and feature detection and 3D calculation. In the camera calibration, it is the way between traditional calibration and self-calibration——Zhengyou Zhang's camera calibration method that is based on plane calibration cyclostyle is adopted. Calibrating template corners' image coordinate are obtained by handwork and Harris corner point detecting, afterward get precise camera parameters by experiment and calculating. Stereo matching is the search of the imaging position of the same target point in two images with binocular vision, and the selection of match features, the definition of match rules and the implementation of match arithmetic are the three important steps in stereo matching. As to the object with circular character, enhanced Hough transformation is applied to accurately extract object's centre of the circle, and then feature matching of left image and right image is carried out. In the feature-based stereo matching method, the SIFT feature extraction and matching algorithm is focused on to obtain the image coordinates of matching points. Finally, according to the basic principle of stereo vision and the computational method of 3D coordinate, the 3D coordinates is obtained. Comparison to the measurement results of the laser sensor, the errors of the obtained distance information of the characteristic points are achieved, and the cause of the error is analyzed. Finally the depth map of 3D environment is produced with the measured results.VC++.NET and OpenCV are used to compile the arithmetics for the experiment in the thesis.