A Biologically Inspired Bayesian Model of Visual Attention for Humanoid Robots

Abstract

Visual attention of human being refers to the process of selectively choosing a set of relevant visual information for further processing. Human eyes are always exposed to enormous amount of visual information, not all of which are relevant to the current mental/behavioral state. The attention system helps to focus on a relevant region of a scene or an object of interest and ease the information processing. This paper proposes a novel probabilistic model of visual attention using recursive Bayes filter and Gaussian adaptive resonance theory (ART) while mimicking some of the major aspects of primates visual attention system (albeit reduced complexity). The target application of the proposed model is intelligent autonomous agent namely, a humanoid robot. The proposed model adopts the propositions of 'biased competitive hypothesis'. Such a model will allow a humanoid robot to autonomously engage its attention to perceptually interesting and/or behaviorally relevant stimuli and learn about it. Preliminary experimental results demonstrate different aspects of the proposed visual attention model