抄録
In this paper, we propose to develop a modular real-time robot vision system in a programming environment based on BeNet. BeNet is a parallel computational model for describing real-time systems completely and briefly by software as a set of computational modules of which roles are distinct. The real-time behavior of BeNet does not depends on the hardware and is easy to predict because each module changes the state and the output at regular intervals. An environment in which one can describe a BeNet and implement it immediately on a multi-processor system with image frame memories, makes it possible to develop a large-scale modular system by integrating diverse methods for computer vision in a short term. We present how to design and develop a system that can find an object and track it in real-time based on the top down input of the attention in our environment. We realized a network of modules for feature extraction, object extraction and attention control, and confirmed that the BeNet shows a desirable behavior in the real world and that it is applicable to an autonomous robot.
