Object Detection with Spiking Neural Networks on Automotive Event Data

Description

Automotive embedded algorithms have very high constraints in terms of latency, accuracy and power consumption. Event cameras and spiking neural networks can be used to address the intricate challenges associated with object detection in automotive applications,emphasizing the critical demands of low latency, high accuracy, and minimal power consumption. Event cameras, with their characteristics of low latency, high dynamic range, and low power requirements, record binary event data that encodes changes in brightness with spatial, temporal, and polarity information. However, the asynchronous and sparse nature of event data poses problems for conventional neural networks designed for processing frame-based data. Spiking neural networks, designed to communicate throughdiscrete binary spikes, are uniquely suited to the inherent characteristics of event data. The specific operation of spiking neural networks, exhibiting a remarkable energy efficiency advantage, is estimated to be 5 to 8 times more efficient compared to traditional neural networks, all while maintaining levels of accuracy. The research harnesses advancements in spike backpropagation and training algorithms to enable our experimental results on real-world automotive event datasets, demonstrating significant advancements in classification performance and thereby achieving state-of-the-art in spiking neural networks. Furthermore, the culmination of this research introduces apioneering development, the integration of spiking neural networks with Single Shot MultiBox Detector (SSD), marking a significant achievement as the first spiking neural networks capable of object detection on the intricate GEN1 Automotive Detection Event dataset.

Abstract

National audience

Additional details