Development of a biologically inspired real-time visual attention system

Abstract

The aim of this paper is to present our attempt in creating a visual system for a humanoid robot, which can intervene in non-specific tasks in real-time. Due to the generic aspects of our goal, our models are based around human architecture. Such approaches have usually been contradictory, with the efficient implementation of real systems and its demanding computational cost. We show that by using PredN1, a system for developing distributed real-time robotic applications, we are able to build a real-time scalable visual attention system. It is easy to change the structure of the system, or the hardware in order to investigate new models. In our presentation, we will also present our system with a number of human visual attributes, such as: log-polar retino-cortical mapping, banks of oriented filters providing a generic signature of any object in an image. Additionally, a visual attention mechanism -a psychophysical model -FeatureGate, is used in eliciting a fixation point. The system runs at frame rate, allowing interaction of same time scale as humans.