Issues, Architectures and Techniques in Real-Time Vision

Abstract

We discuss some issues in real-time vision for robots, presenting an integrated collection of techniques, namely model-based prediction, speculative computing, foveation, and multiresolution processing. We describe an asynchronous, parallel distributed architecture of autonomous agents to support these techniques, in which the agents can be regarded as expert translators between languages at different levels of representation. These techniques have been implemented in several partial, prototype systems, which demonstrate quite impressive real-time performance using only limited computing resources.