|
In this paper we suggest a computational model for the 3D interpretation of a 2D view based on the selection and organization of useful and discriminatory features for a large number of objects designed on a CAD system. These features are used to construct a strategy hierarchy for recognition, which provides the representation of associations between features detected from multiobject scenes and the data base of object models, enabling the on-line recognition to be particularly efficient. The proposed model has been implemented in a model-based computer vision system that can recognize threedimensional objects from unknown viewpoints in single gray-scale images. The system is based on an off-line model preprocessing stage and an on-line recognition stage. In the off-line processing, 3D recognition oriented models, which are used in the verification process, and the strategy hierarchy, which is used in the matching process, are automatically determined.
During the on-line recognition, three steps are performed: feature extraction, matching and verification. The process of feature extraction first applies traditional low-level image processing to detect feature primitives and then produces a description in terms of contour features and relationships between them, based on the statements made by the phenomenon of Perceptual Organization, as collinearity, parallelism, connectivity, and repetitive patterns among image elements. Matches are then made on this intermediate representation, the feature groupings. This process generates initial hypotheses for objects and viewpoints by accessing the strategy hierarchy constructed in the off-line stage, integrating both top-down and bottom-up approaches. Finally, the verification phase solves the spatial correspondence, verifying whether the match leads to a legal interpretation of the image.
|
