文献类型：期刊文献

英文题名：PE-Fusion: A Lightweight Traffic State Recognition Framework Based on Physics-Entropy Fusion for Edge Devices

作者：Zhang, Peng[1]; Wang, Yufei[1]; Yu, Bowen[1]; Yuan, Songjian[1]; Wang, Kai[1]; Rao, Zhiqiang[1]

机构：[1] College of Urban Rail Transit and Logistics, Beijing Union University, Beijing, 100101, China

第一机构：北京联合大学城市轨道交通与物流学院

年份：2026

起止页码：2503-2509

外文期刊名：2026 9th International Conference on Advanced Algorithms and Control Engineering, ICAACE 2026

收录：EI(收录号：20262420910277)

语种：英文

外文关键词：Anomaly detection - Arts computing - Computer graphics - Computer hardware - Convolutional neural networks - Edge computing - Edge detection - Graphics processing unit - Highway accidents - Highway engineering - Intelligent systems - Intelligent vehicle highway systems - Motor transportation - Program processors - Real time systems - State estimation - Traffic congestion

摘要：The exponential growth of urban road networks has precipitated an urgent demand for intelligent monitoring systems capable of real-time classification of diverse traffic states, including congestion, accidents, and fire incidents. However, the deployment of state-of-the-art Deep Learning (DL) models, such as Convolutional Neural Networks (CNNs) and Vision Transformers, is severely impeded by their heavy reliance on Graphics Processing Units (GPUs), rendering them unsuitable for resource-constrained edge devices widely used in municipal infrastructure. To bridge this gap, we propose PE-Fusion, a lightweight framework that synergizes traffic flow theory with computer vision. Unlike purely data-driven approaches, PEFusion extracts kinetic energy fields via sparse optical flow and computes spatiotemporal motion entropy to quantify traffic disorder. By integrating these features with a visual density estimator grounded in the Greenshields model, we achieve precise multi-class state recognition. Experimental results on UCSD and simulated datasets demonstrate that PE-Fusion achieves a mean Average Precision (mAP) of 98.8% and operates at 55 FPS on non-GPU hardware. This performance outperforms YOLOv5-Nano by 6.8 × in speed while maintaining superior accuracy in complex accident scenarios, validating its potential for city-scale deployment. ? 2026 IEEE.