An efficient method for recognising polyhedral scenes using range data

The recognition of three dimensional (3D) shape from range data is an important problem in machine vision. A model-based vision system for recognising polyhedral scenes is presented in this paper. The system uses a single range image view to identify and to locate polyhedral objects in a scene, assuming that objects may be viewed from any direction and that they may be partially occluded by other objects. The proposed system consists of two stages. A segmentation stage that is based on the Laplacian operator which provides an edge map. A recognition stage in which connected components (local patterns) representing visible parts of objects in the scene are matched against the basic polyhedral primitives of the models database. Then, the model objects which are semantically described using basic polyhedral models are matched against the scene components already recognised in the previous stage. The proposed system has been tested on a large number of polyhedral objects in arbitrary views. Initial results are very promising with regard to the system reliability and the speed of segmentation and matching processes