Real-Time Pedestrian Detection at the Edge on a Fully Asynchronous Neuromorphic System

Object detection is a classic application in machine-learning where artificial neural networks have become the most dominant types of model used. Their increasing performances in the last few years also bring an ever-growing computational cost that is incompatible with the limitations of embedded systems at the edge. By taking inspiration from biological systems that have been naturally selected for their efficiency, neuromorphic computing aims at proposing energy-efficient alternatives that are better suited for these types of applications. For this reason, inventions like event cameras and spiking neural networks have been steadily growing in popularity in recent years. To the best of our knowledge, we propose the first fully asynchronous, end-to-end neuromorphic system performing real-time object detection focusing on pedestrians using Spiking Neural Network (SNN) implemented on specialized hardware. Based on previous analysis on the importance of the size and number of events constituting objects in event-based dataset, we also propose a novel weighted loss function that prioritizes the learning of features on eventful object.