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 consists 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 paper presents a new model-based object recognition scheme for general two dimensional objects in a cluttered scene. The scheme considers objects subjected to similarity transformations (i.e., a combination of rotation, scaling and translation). It employs a new feature detection algorithm, combining curvature measures and polygonal approximation. An approximate, but efficient matching strategy is proposed for hypothesis generation and synthetic verification procedures are introduced to improve the robustness of the system. Experiment results are presented to show that the system works effectively and efficiently.