Attention Neural Networks for Pan-Tilt-Zoom Control with Active Hand-Off

Abstract

Communities of cooperating robots would be highly advantaged by the ability to focus the attention of better placed robots upon activities tagged as important by other robots. Neural network and deep learning methods are increasingly applied to attention based steering of cameras and other sensor arrays resident on robots. a hand-off of focus of attention requires that one robot communicate to other robots system state information. The specific state information that needs to be communicated can be difficult to determine in many empirically tuned neural deep learning systems. In this paper, we will propose a method for cleanly transferring focus of attention across physically disjoint deep network based motion trackers. The method has been constructed to have explicit and understandable hand-off capabilities to support tracking of an object of interest across an array of sensors each resident on a disjoint robot or other autonomous agent acting as a community. We will additionally provide an experimental analysis of system efficacy and a discussion of possible future work and the long-term implications of the observed results.