Model-Based Silhouette Recognition

摘要

The authors present a system for recognizing 3-D objects at unknown orientations from their 2-D silhouettes. The geometric description of an object model is provided in CAD form and is then compiled into a set of geometric constraints for a large set of viewing directions. The silhouette is parsed into a set of straight edges, and these edges are compared to the edges of the model by conceptually structuring all possible interpretations in a tree. This enormous search space is pruned by extending the interpretation tree search of Grimson and Lozano-Perez to work for the 3-D model/2-D data case. This includes a precise analysis of the propagation of errors in the position and orientation of silhouette edges, which then provide adequate constraints for pruning the search tree. Any hypotheses that survive the pairwise constraints of tree search are verified by synthesizing a silhouette of the model for the hypothesized orientation and comparing this synthetic silhouette to the observed silhouette. Based only on silhouette data, the system can find all plausible interpretations of the data, including symmetric viewpoints. The system performs in the presence of unknown viewpoint, moderate scale uncertainties, occluding objects, and degradations in the silhouette shape due to image noise and image processing artifacts. These characteristics should enable the system to perform well in applications where images have reasonable spatial resolution but where limited resolution in the signal (intensity or range) reduces the information in the data to a silhouette. Keywords: Computer vision. (KR)