Neuromorphic Event-based Line Detection on SpiNNaker

The combined use of Spiking Neural Networks (SNNs) and neuromorphic data in recent years makes for a promising solution to the challenges currently raised in computer vision. Indeed, the natural match between SNNs and event data leads to improvements in terms of biological inspiration, energy savings, latency and memory use for dynamic visual data processing, especially when such networks are implemented on neuromorphic hardware. We propose to draw advantages from these technologies to propose, to the best of our knowledge, the first end-to-end neuromorphic model for straight line detection, a standard task in robotics and computer vision. Our architecture relies on SNN intrinsic dynamics and ensures the accurate detection of moving lines recorded by an event-based camera with no learning. It reaches an overall performance of over 90 % with a limited number of neurons and synapses allowing for its deployment on the neuromorphic board SpiNNaker.