How the Brain Achieves Real-Time Vision: A Spiking Position Perception Model

摘要

Real-time visual perception is essential for animals to survive in a complicated natural environment and for robots to interact with moving targets. However, delays generally occur during the signal transfer and processing both for animals and robots, and these delays would produce errors during real-time interactions with the physical world. Natural facts have shown that animals can perfectly compensate for these pervasive delays. In this paper, we propose a novel and effective position perception model (PPM) based on spiking neural networks (SNNs) to address this ambivalent situation in robotic vision systems. We investigate the performance of PPM by tracking a moving target. PPM can compensate for temporal delays in the system regardless of the target’s speed. We also present a deep version of PPM (dPPM). dPPM can handle some more complex situations and make long-term anticipations. We finally implement PPM on neuromorphic chips and test it on real DVS (dynamic vision sensor) data, and it can perform real-time or anticipative visual perceptions.