A Real-Time Event-Based Selective Attention System for Active Vision

In real world scenarios, guiding vision to focus on salient parts of the visual space is a computationally demanding tasks. Selective attention is a biologically inspired strategy to cope with this problem, that can be used in engineered systems with limited resources. In active vision systems however, the stringent realtime requirements limit the space of solutions that can be achieved with conventional machine vision techniques and systems. We propose a hybrid approach where we combine a custom neuromorphic VLSI saliency-map based attention system with a conventional machine vision system, to implement both fast contrast-based saccadic eye movements in parallel with conventional visual attention models that use high-resolution color input images. We describe the system and characterize its response properties with experiments using both basic control visual stimuli and natural scenes.