| The image based modeling technology has drawn the interests of many researchers in recent years. Observe the same object from two or more viewing points to obtain many images, remove the noise by image preprocessing, and then acquire the disparity of correspondence pixels, compute the depth information of the object in scene using the disparity , finally, we could achieve the spatial position of the object. In the practical application, binocular stereo vision is closer to the actual people's binocular version principle, and is easy to be implemented, this thesis analyzes and researches the 3D reconstruction about binocular stereo vision. While compute the 3D information using the disparity, stereo matching is the key and difficult problem in research of computer 3D modeling.,which can be divided into point matching, line matching, area matching, and which also can be divided into feature matching, region matching, phase matching and energy matching. Stereo matching is an important research subject in computer vision and pattern recognition, whose application has infiltrated into object recognition, content-based image retrieval, image mosaic, vision measurement and other fields. Most of stereo matching theories and technologies are closely related to practice. There are almost no general theories and technologies, which has not been thoroughly solved up to now. Aiming at practical application, most researchers now study on different matching algorithms to provide the data in 3D reconstruction. This thesis concentrates on the research on some key techniques of stereo vision, such as camera calibration, corner detection, stereo matching and 3D reconstruction.The main progress in this thesis is as follows:(1)In region matching, a new method is proposed. Firstly, the images would be decomposited usingàtrous wavelet. Secondly, the different sub-bands of left and right images are computed by causal neighborhood hierarchical overlapped block method, At last ,the dense disparity map could be obtained.(2)In order to improve the stability of feature extracting and matching, this article propose a method to select multi-scale Harris features through building scale-space, and describe the feature with a stable descriptor.This method can extract and match features well in every condition.(3)In camera calibration ,the paper discusses a new linear approach. A reference object is visual in both images. Obtain the necessary information of known points ,and then compute the relative position matrix of left and right cameras.(4)The 3D space information calculation. Combining with matching point coordinates and camera matrix, using dependent derived formula, we could calculate the 3D coordinates by least squares method . (5)About 3D modeling ,we adopt two methods: First, the AutoCAD and 3Dmax are used to model and texture. Second, we program with OpenGL to Model.The analysis of the algorithm and experimental results show that:(1)The region matching algorithm is effective ,and the accurate matching point pairs and dense disparity map could be obtained. The disparity could be applied to 3D reconstruction.(2)The feature descriptor about feature extracting and matching is of shift, scale, illumination and rotation invariance. It's fixed for different condition(3)The camera calibration based on reference object possesses a certain accuracy. The space coordinates of 3D reconstruction reflect the space information of the target objects, which could be displayed vividly. |
PDF Full Text Request