Optimized feature exploitation for 3D object recognition using ART neural networks

In this paper, a study is presented of how self-organizing ART networks can be used to create a trainable, feature-based real-time 3-D object recognition system. Feature extraction is a well known approach to reduce the number of appearances of a three-dimensional object. Since features are derived from only a small part of the information comprised in the original image, it cannot be assumed that a given set of objects is separable in the reduced feature space. To avoid ambiguities, in general, multiple features have to be integrated in an object recognition system. Since feature extraction can be computationally intensive, a real-time system should evaluate features sequentially and terminate recognition when ambiguities are resolved. This paper gives an analysis of the clustering properties of ART 2A-E networks. It is shown how ART networks can be used to generate meaningful hints concerning the object's identity from ambiguous features by exploiting them up to an optimal degree.