The most important issues in computer vision are visual metrology and pose determination. Although fruitful achievements have been obtained in the field, there still exist many difficult problems to hamper the real progress of applications. This thesis is focused on some real problems about visual metrology and pose determination. The main work can be summarized as:1)It is a conventional belief that line-based approaches perform better than point-based ones for homography estimation, as the line-fitting is generally more noise resistant than point detection. In this work, we show that blithely using line-based estimation is a risky business. More specifically, we show that when the image line(s) is (are) passing through or close to the origin, the line-based homography estimation could become wildly unstable whereas the point-based estimation performs normally. To tackle this problem, a new normalized method specially designed for line-based homography estimation is proposed and validated by extensive experiments.2)An automatic approach for detecting the contacting point of wheel to the ground is proposed. The fundamental principle of this approach is the fact that the contacting point must lie on the line going through the projection of the wheel center and the vertical vanishing point. Extensive experimental results show that the proposed approach is capable of detecting the contacting point satisfactorily and effectively.3)A simple and efficient method based on planar homography is proposed for face pose estimation. The method is composed of the following steps: At first, the parameters of the 3D planar face model are estimated from a single fronto-parallel face image. Then the homography between the face model plane and the image plane is estimated, by which the face pose and the face image location are determined. Finally, the M-Estimator is employed to further refine the results. The main characteristic of the proposed method is that it is simple to implement, does not need to calibrate the projective camera beforehand and can recover the face pose and the face image location satisfactorily under a wide degree of head motion.4)Two novel robot self-localization methods based on planar laser measurement are proposed. The first one is an improved Hough density spectrum based method, where a novel Hough density spectrum is introduced, and by which the location accuracy and robustness are both enhanced. The second one is a Fourier-Mellin transform based method. In this method, by firstly converting the two measurement point sets into two binary images, the Fourier-Mellin based image registration technique, a popular method in image registration field, is employed to determine the rotation parameter at first, then followed by a standard ICP technique with the Hausdorff distance as its cost for the estimation of the two translation parameters. The experimental results show that both our proposed methods can perform robustly and accurately.
修改评论