GPU-accelerated affordance cueing based on visual attention

摘要

This work focuses on the relevance of visual attention in affordance-inspired robotics. Among all approaches in robotics related to Gibson's concept of affordances the dealing with attention cues is only rudimentary. We are introducing this concept within the perception layer of our affordance-inspired robotic framework. In this context we present a high-performance visual attention system handling invariants in the optical array. This layer builds the base of higher-sophisticated tasks, like a "curiosity drive" that helps a robotic agent to explore its environment. Our attention system derived from VOCUS utilizes the parallel design of the graphics processing unit (GPU) and reaches real-time performance for the processing of online video streams in VGA resolution on a single computer platform. GPU-VOCUS is currently the fastest known visual attention system running on standard personal computers.