Recognition and localization of a 3D polyhedral object using a neural network

This paper proposes a centroidal profile (CP) and neural network based 3D object recognition and localization method (PRONET). In PRONET approach, CP patterns are extracted from a multiview model of a 3D CAD representation of an object. Correspondences are also saved in the CP pattern. A three-layer feed-forward neural network is trained with these CPs. By matching a CP pattern of the given image with that of the neural network, approximate orientation of the object and line-to-line correspondence between image features and CAD features are obtained. An iterative model posing method then calculates the more exact pose of the object based on initial orientation and correspondence. The advantages of this method are (1) since time-consuming portions of the task are executed off-line, an object can be quickly recognized in the execution stage, and (2) correspondences between 2D image features and 3D model features can be quickly obtained by matching a single CP pattern, instead of creating many hypotheses and then trying to verify each, and (3) PRONET is tightly integrated with the CAD system.