Saliency based sensor fusion of broadband sound localizer for humanoids

Abstract

This work presents a biologically inspired solution to problems that arise in multisensory attention, with a specific application to binaural humanoid robotics. The focus was on using only two microphones as an analogy to mammalian auditory system. The goal was to localize a salient sound source and to fuse this information with a visual-salience feature selection system. We describe a method to select task-relevant sounds in the auditory scene using both interaural time difference (ITD), inspired by the work of Jeffress and Konishi in the avian auditory system. We modelled the well-studied coincidence detectors of the barn owl as banks of frequency-tuned delay-and-sum beamformers, with frequency decomposition based on a Gammatone model of the human cochlea. Furthermore, we developed a useful metric of auditory salience that emphasized onsets of spectrally complex sounds. Finally, we developed an interface between this auditory-salience based attention orienting system and an existing visual-salience based attention system on the iCub humanoid robot. We demonstrate that the iCub is capable of behaviours that would be impossible without fusion of auditory and visual attention.