A new model-based approach for recognizing occluded curved objects in dense-range images

Although the use of range imaging has been gaining popularity in 3D object recognition, it is still widely accepted that recognition of real world scenes, with possible occlusion, and based on a single view, is a difficult task. In this paper, some of these recognition problems are investigated. In particular, a new model-based matching technique is applied to recognize a scene with many occluded curved objects using a single range image viewed from any direction. The proposed method is capable of recognizing single objects as well as configurations of objects. The segmentation process is based on two newly developed directional curvature operators. The resulting data, expressed in terms of linear segments, cylindrical arcs and spherical arcs, is then transformed into a surface-based representation that is enforced by a boundary representation scheme. Objects in the scene are recognized using a matching strategy based on the generation of relative bases to models (RBM). If a match is found, the model appearance with reference to the RBM will be similar to that of the object in the absolute base. Once all objects have been recognized, the system attempts to match configurations, if there are any. By configuration here we mean, a set of primitive-objects related to each other by spatial relationships. The system has been tested on a large number of real range images and synthetic data. The experimental results have shown that the system can successfully recognize complex scene with an acceptable reliability and accuracy