Rigid-Body 3-D Motion And Structure From Feature Correspondences

This dissertation is the project supported by Aeronautical Foundation Science (No. 03I53062).Visual motion analysis has been a major topic of research in computer vision. It is the method of rigid-body 3-D motion and structure recovery from image sequences. Methods for it can be classifed as correspondence-based and optical flow-based. This dissertation discusses the former only, recovering the rigid-body 3-D motion and structure from features between the image sequences. Once the motion object is detected, computer vision system can recognize it and estimate the current and future object location and pose with the motion and three-dimensional structure information, so it can be used to resolve the problem of object recognition and object tracking. Visual motion analysis is widely used in military, commercial, industrial, medical, meteorological field and so on.The first step for extracting structure and motion information from image sequences is the necessary image pre-processing, image features (feature point and feature line) extracting and image features matching;then establishing a set of equations according to the rigid-body motion and projection equations, and getting the motion and structure parameters. The dissertation introduces the polular methods of feature extracting and matching briefly, while focuses on motion and structure parameters estimation.Previous work in rigid-body motion from feature correspondences focused either on two (or more) perspective images or on a set of orthographic (more generally, affine) images. In practice, a thorough study of structure from motion using combinations of close shot and distant shot is needed. Usually one can use the orthographic model for distant shot, and the perspective model for close shot.Taking the practical application into account, this dissertation presents a first investigation on the structure from motion problem from the combinations of perspective and orthographic images based on feature correspondences. The algorithmsimplified epipolar geometry without affectting the recovery of Euclidian structure. The dissertation introduces the foundational information about 3D motion estimation firstly, then focuses on two cases: 3D motion and structure between one perspective and one orthographic views from point correspondences, 3D motion, and structure between two perspective and one orthographic views from line correspondences. Its contribution includes: description of the epipolar geometry, analysis of the particular fundamental matrix, presentation of simple and effective linear algorithm. Computer simulation with data and artificial images show that the performance of the technique is effective and stable.