| Computer vision is a new research field with fast development into one of the important fields of intelligent automation. Its research object is to equip the computer with the ability to acquire information from its surrounding images or combination of images, which enables the machine not only to sense the geometrical information of the object in the environment, including its shape, position, pose and motion, etc., but to describe, store, recognize and understand them.Firstly, this dissertation introduces the basic conceptions and relevant properties of projective geometry, affine geometry, metric geometry and Euclidean geometry used in computer vision, as well as the camera model with imaging principles. Secondly, the dissertation gives review of the development of camera calibration technique in recent years. Some analysis, comparison and categorization on the existing various camera calibration techniques are discussed. Finally, the focus in my dissertation is the research on the camera self-calibration with varying intrinsic parameters, the perspective correction of plane scene image, derivations of four nonlinear state estimation filters (EKF1, EKF2, DD1 and DD2), and the estimation of pose and motion based on monocular vision. Linear camera self-calibration with varying intrinsic parameters. In some vision systems (e.g., the vision system of robot and active vision system), the camera's position and optic system (for example, the aperture and focus) require constant adjustment, and the camera must recalibration after every adjustment. Aimed at this circumstance, this paper proposes a camera self-calibration method to deal with the situation when the camera's distorted skews and principal points are already known, while the other intrinsic parameters keep varying. That is, to calculate the fundamental matrix between images to obtain projective reconstruction at first, and on which basis then, to regain the homography matrix by linear method, lastly, to obtain the camera's intrinsic parameters using homography matrix. 3D metric reconstruction based on scene geometry. The 3D metric reconstruction of scene image is often aimed to the image sequences, and the stratified metric reconstruction method is to obtain the projective reconstruction of image sequences at first. If the surfaces of 3D object are plane (This situation is familiar), then based on the perspective correction of plane scene image and the scene geometry information, we give a review of the perspective correction of plane sceneimage and the metric reconstruction of routine 3D scene image. Then on the basis of this, this dissertation proposed a 3D metric reconstruction method of single view which don't need to obtain the projective reconstruction. Four kinds of nonlinear state estimation filters (EKFl, EKF2, DD1 and DD2). Some system can't be described by simple linear model, therefore, the development of the non-linear filter algorithm becomes a necessity. Based on Kalman filter, the first order and second order Taylor series can be adopted to approximate the non-linear dynamic equation and measurement equation, and to obtain filters of EKFl (on the basis of first order Taylor approximation) and EKF2 (on the basis of the second order Taylor approximation) respectively. Whereas, the EKFl and EKF2 filters demands the existence of corresponding the first order and the second order differentiation of non-linear dynamic equation and measurement equation. If this requirement is not satisfying, Stirling interpolation approximation which does not need differential calculus, in replacement of Taylor approximation, gives forth to another two filters, DD1 (based on the first order Stirling interpolation approximation) and DD2 (based on the second order Stirling interpolation approximation). Estimation of pose and motion based on monocular vision. With the emergence of the new-born science of computer vision, it is also very important to estimate the correlative three-dimensional pose and motion i...
|
PDF Full Text Request