spike-yolo-automotive-perception

name	spike-yolo-automotive-perception
category	ai_collection
description	First comprehensive evaluation of SNNs for real-world automotive multi-object detection and tracking using SpikeYOLO transfer learning. Achieves mAP 0.937 (KITTI) and 0.771 (BDD100K MOT2020) for detection, HOTA 0.701/0.445 for tracking — competitive with conventional DL, with energy-efficient edge deployment. arXiv:2607.04921
source	arXiv:2607.04921
arxiv_id	2607.04921
trigger_words	["neuromorphic automotive perception","SpikeYOLO","SNN object detection tracking","neuromorphic autonomous vehicles","energy-efficient edge perception","event-based object detection","SNN multi-object tracking","KITTI neuromorphic","BDD100K neuromorphic","carbon footprint deep learning"]
created	2026-07-11
updated	2026-07-11

Efficient Perception in Automotive Detection and Tracking Using Neuromorphic Computing

Paper: "Efficient Perception in Automotive Detection and Tracking Using Neuromorphic Computing" — arXiv:2607.04921 [cs.CV, cs.AI], July 6, 2026

Abstract Summary

Deep learning algorithms are notorious for high carbon footprint and computational demands that limit edge deployment and raise sustainability concerns. Neuromorphic computing and Spiking Neural Networks (SNNs) offer energy-efficient alternatives with massively parallel computation and on-chip learning. This paper presents the first comprehensive evaluation of SNNs for real-world automotive multi-object detection and tracking. Using transfer learning with SpikeYOLO, achieves mAP 0.937 on KITTI and 0.771 on BDD100K MOT2020 for detection, and HOTA 0.701 (KITTI) and 0.445 (BDD100K MOT2020) for tracking — results competitive with conventional deep learning methods.

Key Innovations

1. First SNN Evaluation for Automotive Perception

• Comprehensive study of SNNs for real-world autonomous vehicle perception
• Covers both object detection and multi-object tracking
• Addresses the critical gap: can SNNs compete with conventional DL in safety-critical automotive tasks?

2. SpikeYOLO Transfer Learning

• Adapts the YOLO architecture to spiking neural networks
• Transfer learning from pre-trained ANN to SNN
• Maintains competitive accuracy while enabling energy-efficient inference
• Bridges the gap between ANN performance and SNN efficiency

3. Real-World Benchmark Results

Object Detection:

Dataset	Metric	SpikeYOLO	Conventional DL
KITTI	mAP	0.937	~0.94-0.96
BDD100K MOT2020	mAP	0.771	~0.78-0.82

Object Tracking:

Dataset	Metric	SpikeYOLO	Conventional DL
KITTI	HOTA	0.701	~0.72-0.75
BDD100K MOT2020	HOTA