Model-based Recognition of Occluded Curved Objects

R. Benlamri (United Arab Emirates)


Range image segmentation and representation, Curvature operators, Model-based recognition.


This paper describes a new model-based approach to recognize occluded curved objects from single range images. Image segmentation is based on surface curvature detection. An edge-junction graph and a surface adjacency graph are then used to infer surface patches and their spatial relationship respectively. The matching algorithm is based on Local Geometrical Patterns (LGP's) extracted from adjacent surfaces. The idea is to generate, for each similar pair of object and model LGP, a location and orientation transformation expressed in terms of a Translation Rotation Matrix (TRM), that allows a model (defined in its model coordinate system) to be aligned onto its corresponding object (defined in the viewer centered coordinate system). Unlike existing methods, only one LGP is used to check for matching. Furthermore, once the two coordinate systems are aligned, no more geometrical transformations are involved in the matching process. This significantly enhances the matching search-space and reduces its execution time. The system has been tested on real range images and synthetic data. The experimental results have shown that the system can reliably recognize scenes with acceptable accuracy.

Important Links:

Go Back