On the recognition and location of partially occluded objects

The recognition and location of partially occluded objects is important for image-guided robot automation. A computational object recognition system c onsists of three main parts: shape representation, matching strategies and verification. The shape representation scheme, which is always application- oriented, should keep extracted features as invariant as possible. This pap er presents a new model-based object recognition scheme for general two dim ensional objects in a cluttered scene. The scheme considers objects subject ed to similarity transformations (i.e., a combination of rotation, scaling and translation). It employs a new feature detection algorithm, combining c urvature measures and polygonal approximation. An approximate, but efficien t matching strategy is proposed for hypothesis generation and synthetic ver ification procedures are introduced to improve the robustness of the system . Experiment results are presented to show that the system works effectivel y and efficiently.