Application of Hopfield neural networks and canonical perspectives to recognize and locate partially occluded 3-D objects

Abstract The task of recognizing and locating the partially occluded three-dimensional (3-D) rigid objects of a given scene is considered. The surfaces of 3-D objects may be planar or curved. The 3-D surface informations are captured through range data (depth) map. For recognition we use the principal curvatures, mean curvature and Gaussian curvature as the local descriptions of the surfaces. These curvatures do not change significantly under rotation and translation. Hence they are used as local invariant (within certain threshold) features of the surfaces. A neuro-vision scheme, based upon the matching between the local features of the 3-D objects in a scence and those of the object models is proposed. Object models are generated using canonical perspectives. The feature matching scheme is realized, at two stages, through Hopfield neural networks. At the first stage edge-points of the scene are matched with those of the object models using a Hopfield net. At the second stage non-edge-points of the scene are matched with those of the object models using another Hopfield net. Compared with conventional object matching schemes, the proposed technique provides a more general and compact formulation of the problem and a solution more suitable for parallel implementation. Finally, the hypothesis generation and verification scheme is proposed for best possible recognition.