Spatial contours for vision and CAD model matching

摘要

Off-line programming techniques for artificial vision systems attempt to exploit existing databases. Sparing the reference image of a real part, the learning process is essentially based on 3D geometrical models, little adapted to image analysis. We suggest an intermediate description using spatial outlines, including global and local data. The former eliminate a large number of potential cases and the latter ensure the validity of the remaining solutions. This is a composite approach for the pattern matching process. The learning phase supplies us with a vision-oriented description obtained from a CAD model. Inversely, during the recognition phase, a CAD-oriented description is extracted from an image sequence. Tested on polyhedric samples and applied in the robotic field, this description has proved to be simple, robust and precise for part identification and localization using monocular vision.