Image-Based Modeling and Rendering is a new method of generating realistic scene view. IBMR integrates computer vision and traditional computer graphics, and relates with stereo vision closely. This thesis focuses on the rendering realistic scene view based on stereo vision.First, this thesis studies the theory and technology in stereo vision, and the methods of camera self-calibration especially. This thesis develops an auto and precise method of camera self-calibration. That is, the feature points-corners are matched and rectified after they are detected, then the precise fundamental matrix is estimated by the improved Bayes weight method. After that the matching relation of images is established.Second, the image matching is realized based on wavelet transform. At first the relationship between wavelet transform and stereo matching is analyze by example, then, the similar distance is defined to measure the similar degree of correspondence points during the disparity estimate. During the stereo matching, epipole constrain is used to find correct match points. At last, the disparity map describes the stereo matching of images points.Finally, the rendering methods are presented in this thesis. One of them is to recover 3D structure of scene by fundamental matrix. The other is rendering new image by view morphing technology.Through the all processes, the methods of 3D reconstruction and rendering based on stereo vision are presented by theories and experiments.
|