Perception of shape and motion

This thesis is concerned with computational fundamentals and robust algorithms for the problem of estimating three-dimensional motion and structure from image sequences under either perspective or orthographic projection. A theoretical framework and a system of methods are developed in this research. The system of methods includes formulations and algorithms for feature extraction, matching, motion estimation, and surface reconstruction.New algorithms for detecting point features and edges are developed in this research. These algorithms improve the existing ones in that the new ones use adaptive thresholds and have better localization accuracies. The detected features and edges are quantized for further use such as matching.Robust algorithms for obtaining unambiguous matches of point features and edges are developed, which enforce intensity or contour consistency, and geometric, rigidity, disparity, and motion epipolar (if available) constraints. The algorithms are so designed that only under rare situations, which occur with zero probability, mismatches can be admitted in occluded regions.Uniqueness conditions for determining structure and motion from monocular image sequences under either perspective or orthographic projections are established. Robust algorithms for estimating structure and motion parameters from correspondences or trajectories of points or lines are developed, and good results with real image data have been obtained.A novel algorithm for reconstructing surface shape from only matched edges and features is developed, which correctly finds depth discontinuity boundaries and occluded regions without using surface models.Physics-based vision problems such as optical flow and shape from shading are also addressed. New concepts and methods such as generalized, 3D, constrained, and parametric optical flows are introduced which would allow more robust and realistic solution of the optical flow problem. Ruling effects of contour, surface marking, and background are discussed which show the fundamental flaws of the existing shape from shading methods based on differential equations. Synthetic images are generated which show that intensity variation plays only a weak role in human perception of shape from shading while contours and surface markings place more reliable and restrictive constraints on possible surface interpretations. A solution of the general shape and motion estimation problem integrating feature matching, motion-stereo, optical flow, and shape from shading is proposed.A fully automated system for motion and structure estimation that integrates methods for feature detection, matching, motion estimation, and surface reconstruction is presented, which has the capacity to incorporate all visual cues and methods related to shape and motion understanding.U of I OnlyETDs are only available to UIUC Users without author permissio