Hybrid object models for robot vision

摘要

This paper concentrates on object models for the recognition of complex three-dimensional objects with a robot vision system. After giving a short overview on existing approaches, some demands on object models for robot vision systems are formulated. Afterwards, an approach of hybrid object models that fulfils all of these demands is presented. These hybrid models integrate neurobiologically motivated object representations by model neurons similar to complex cortical cells and the explicit representation of objects by semantic networks, a well known methodology in the field of symbolic artificial intelligence. Thereby, one can combine the attribute of robustness and fault tolerance of neural networks with the well structured design of symbolic processing. Additionally, the procedural interface of semantic networks allows the development of active vision systems and the implementation of reliable recognition on the basis of multiple viewpoints.