Leveraging graph attention and temporal fusion for lane-level traffic anomaly detection in vehicle-road collaborative systems

Abstract

As urban traffic systems grow increasingly complex, traditional methods for managing congestion and accidents are proving insufficient. This research proposes a lane-level traffic anomaly detection framework that integrates Graph Attention Networks (GAT) with temporal fusion encoding to capture spatio-temporal dependencies within vehicle-road collaborative systems. A self-supervised learning mechanism enables effective anomaly detection with minimal labeled data while an attention-based module prioritizes critical neighbour regions to enhance robustness and computational efficiency. Through comprehensive experimentation, the proposed framework demonstrates strong anomaly detection performance with an F1-score of 0.84, significantly outperforming established base models such as LSTM (0.74), Autoencoder (0.65), DBSCAN (0.46) etc. This enables more adaptive, efficient and intelligent traffic management while optimizing overall network flow.