Estimation of three-dimensional motion and orientation of rigid objects from an image sequence: A region correspondence approach
Date of Award
Doctor of Philosophy (Ph.D.)
Electrical and Computer Engineering
First Committee Member
Tzay Y. Young - Committee Chair
This research deals with the computer vision task of dynamic scene analysis. The objective is to extract three-dimensional (3-D) motion parameters of a rigid object from an image sequence recorded by a video camera. The motion that we are concerned with is the relative motion between the object and the viewing system. Both the position and the orientation of the object are unknown; however, the object is assumed to have at least one visible planar face. No further knowledge about the object is required.An image is perceived as a collection of regions called segments, where each segment represents a certain logical entity present in the object space. When an object undergoes a motion, it is clear that the segments corresponding to different faces of the object undergo certain shape changes. The current research establishes the fact that one can extract the 3-D motion parameters from the quantitative description of the shape changes of individual regions. Several temporal and spatial configurations as well as imaging geometries are considered to examine the extent of information that can be extracted.In the case of orthographically projected images, it is shown that the shape transformation can be described as a 2-D affine transformation, which is governed by the 3-D rotation parameters only. In contrast, for perspective projection, the shape transformation is proven to be a nonlinear (quadratic) transformation governed by the 3-D rotation as well as the 3-D translation parameters.A set of methods for computing these shape transformation coefficients is developed. First, an iterative method, based on operator formulations is considered. The second approach is based on the moment invariants. The third method, an accelerated iterative technique, combines the above two with reasonably chosen heuristics to improve the convergence. In addition, certain mutual constraints among these coefficients are derived, which play a significant role in shape recovery.It is apparent that the method requires that each image in the image sequence be divided into distinct segments. Consequently, the success of the overall system depends on the robustness of the segmentation system as well. A multistage segmentation system has been developed for this purpose. In addition, the proposed methods are applied to several real world and simulated images to demonstrate the strengths of the proposed approaches.
Gunasekaran, Seetharaman, "Estimation of three-dimensional motion and orientation of rigid objects from an image sequence: A region correspondence approach" (1988). Dissertations from ProQuest. 2701.