Structure and motion estimation and recognition for curved three-dimensional objects

The main focus of research in computer vision in the past has been on polyhedral objects that give rise to viewpoint-independent edges in the image. This thesis focuses on smooth curved objects that give rise to viewpoint-dependent edges. The research presented primarily focuses on the problems of structure and motion estimation and object recognition.In the first part of the thesis, the structure and motion of a smooth object are estimated from its silhouettes observed by a trinocular stereo rig over time. First, a model is constructed for the local structure along the silhouette for each frame in the temporal sequence. The local models are then integrated into a global surface description by estimating the motion between successive frames. The algorithm tracks certain surface features (parabolic points) and image features (silhouette inflections and frontier points) that are used to bootstrap the motion estimation process. Points on the entire silhouette, along with the reconstructed local structures, are then used to refine the initial motion estimate. The proposed approach is implemented and results obtained on real images are presented.The second part of the thesis presents a new algorithm for recognition of curved 3D objects from 2D images. The algorithm is based on a representation composed of a discrete set of HOT curves at which the surface admits High Order Tangents. A method is presented to automatically construct two of the HOT curves (the parabolic and limiting bitangent curves) using the results of the above structure and motion estimation algorithm. There is a natural correspondence between these two HOT curves and certain silhouette features: the inflections and the bitangents. The recognition approach uses these silhouette features to compute a set of scale-independent image observables that serve as indices in a database of models. This database is used for pose estimation and model identification. Hypotheses formed through indexing are verified through pose estimation. The results obtained on real images are presented.U of I OnlyETDs are only available to UIUC Users without author permissio