Toward Knowledge Propagation in an Omnidirectional Distributed Vision System

This paper explores a methodology of knowledge propagation between the Omnidirectional Visual Agents of a Distributed Vision System. The application we chose in this work is to navigate a mobile robot in a real-world environment using a network of uncalibrated cameras. A practical application could be to use an existing camera network intended for surveillance for navigating a service robot in a large scale environment composed of several separated rooms. In this first work, we realised a Vision Agent that learns how to control a mobile robot and propagates this knowledge through the network. The Vision Agent autonomously generates the set of examples used in the learning stage. The knowledge is stored in a neural network, called the learner. Once the Vision Agents has learnt how to control the robot, this knowledge is transmitted to the other Vision Agents in the network. The Vision Agents in the network are heterogeneous, so the knowledge acquired by one agent cannot be used directly by another agent and a certain amount of re-learning is needed. The relearning process is guided by a second neural net, called supervisor. We are more interested in the knowledge propagation process than in the knowledge acquisition, we did not focused in the implementation of the neural nets they were used only as a tool. Preliminary results of the experiments are presented.