Wang, Ziwei (Southeast University), Liu, Zhichao (Southeast University), Wu, Suzheng (Southeast University), Wang, Fanxun (Southeast University), Yin, Guodong (Southeast University)

High-Performance and Energy-Efficient Object Detection in Autonomous Driving with Spiking Neural Networks

Scheduled for presentation during the Regular Session "S04b-Intelligent Perception and Detection Technologies for Connected Mobility" (WE-EA-T4), Wednesday, November 19, 2025, 14:30−14:50, Surfers Paradise 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 19, 2025

Abstract

Object detection plays a critical role in autonomous driving within intelligent transportation systems(ITS), where the complexity of traffic scenes poses challenges for camera-based object detection systems. Besides performance, efficiency also stands as a pivotal factor in autonomous perception system design. Brain-inspired spiking neural networks (SNNs) offer a low-power advantage over traditional artificial neural networks (ANNs), but their application in traffic object detection remains underexplored. In this work a novel SNN-based framework for high-performance and energy-efficient object detection in autonomous driving is proposed. Building upon the SpikeYOLO architecture, new spike-driven feature extraction modules, Spike Convolution Blocks, are proposed to capture features of traffic objects from both shallow and deep layers. A multi-scale and multi-class detection design is introduced to address the inherent scale variation and categorical diversity of traffic objects. Experiment results on Cityscapes and TT100K show that the proposed algorithm improves mAP@50:95 by 10.4% and 12.0% respectively compared to the state-of-the-art SNN method. When benchmarked against representative ANNs, the proposed solution demonstrates competitive detection accuracy while improving energy efficiency by 2.47× and 3.30× on the two datasets, respectively.