Novel frameworks for deformable model and nonrigid motion analysis

It is common to perform a two-step process for the analysis of nonrigid motion in computer vision. The first step is the shape recovery from images. In the second step, shape-based motion analysis and point correspondence recovery is performed on reconstructed object shapes to provide further understanding of the object motion.; This dissertation addresses problems involved in the two-step motion analysis process. The main application area of techniques presented in this dissertation is analysis of medical object motion, especially the motion of human tongue.; Ultrasound imaging is the main approach for tongue image data acquisition, owing to its real-time frame rate and non-invasiveness. We note that to-date, there is no computer vision method to extract band-shaped objects from imagery. The human tongue in ultrasound imagery is band-shaped, and objects such as face in a video can be better tracked by using the "band structure" along the boundary. As a part of our shape recovery techniques, we introduce a novel parametric deformable model suitable for edge extraction. A contour extraction subsystem, EdgeTrak, has been developed based on our parametric deformable model and is currently used in several institutions for studying various aspects of tongue. We also introduce a geometric deformable model which is designed particularly for open contour extraction and end point detection.; In order to obtain the best representation of the human tongue motion, a point correspondence recovery technique with local searching and a shape-based dynamic programming technique for contour time-alignment has been developed. We also introduce a general framework for 2D multiframe and 3D surface-to-surface motion analysis. Techniques introduced in this dissertation have been tested on synthetic and real world data. Experiment results show that our techniques are effective and applicable for practical use.