Some Crucial Technique Research For Large-Scale Surface 3D Measurement Based On Computer Vision

In automotive industry, there is a need for accurately measuring the 3-D shapes of die and stamped parts of the large-scale complex car body to speed up product development and ensure manufacturing quality. 3-D measurement systems based on optics and computer vision can provide a non-contact, rapid and reliable measurement compared to conventional 3-D coordinate measurement machines. It is valuable to research vision measurement technology.Based on analyzing the existing research works and literatures of 3D vision measurement, aimed at the existing problems, some crucial technology for large-scale surface 3D measurement based on optics, digital image processing and computer vision are theoretically and experimentally studied in this dissertation. Camera model is established above all. Some methods of camera calibration, the epipolar geometry constrain and various computational techniques of fundamental matrix are systematically analyzed and compared in this thesis. We present a camera calibration algorithm by viewing a plane using reprojection.A novel approach of large-scale surface 3D measurement is proposed which combine multiple view motion vision measurement, digital fringe projection measurement and iterative closest point registration. The measurement techniques is rapid, portable, easy to operate and large measurement area. Compared with the existing methods, the accuracy of surface merged is not dependent on high accuracy of feature point measurement, but determined by iterative closest point registration.We have studied a new algorithm of feature point correspondence based on artificial landmark, solved the problem of feature point match between images. Artificial feature landmark is first leaded into vision computation in this paper. Our approach is to extract marker points and to use correlation to find an initial set of matches, then use relaxation, robust recover epipolar geometry and reprojection to discard false matches in this set and find more matches. The algorithm is rapid, accurate and no false match points in the main.Feature extraction algorithm, one of the most important areas in computer vision, is studied and algorithms of stereo and closed range photogrammetry are analyzed. This thesis proposes an algorithm of feature point 3D vision measurement from motion to use cycle computation and discard 2D & 3D outliers for the first time.Experiment state clearly the effects of outliers can be removed by the algorithm which is robust and better results have been obtained.Finally, a new 3D surface ranging method based on a digital fringe projection and shifting technique is systematically studied and non-linear rectification algorithms which reduce measurement error is proposed. At the same time, the noise-immune phase unwrapping algorithm is deeply studied and a new algorithm for phase unwrapping of phase map is proposed. Experiment make known the algorithm can reduce measurement error, bypass the noise points automatically, solve the problem of the shade sheltering or cavity, overcome the error propagation problem and better results have been obtained.Owing to 3D vision measurement relating to many intersect disciplines, a lot of technique questions exist still. This thesis only involves some crucial techniques in this field, we make further study to some questions left.