Utilizing Graph Neural Networks to Detect Abnormal Driving Patterns in Vehicular Networks

Abnormal driving presents a significant risk to road safety, impacting both drivers and pedestrians. Meanwhile, vehicular networking has emerged as a crucial technology for enhancing road safety by improving vehicle-to-infrastructure communication. This study leverages vehicular network communications to identify dangerous driving behaviors, such as aggressive or impaired driving. We employ graph-based methods to analyze traffic patterns via network topology, known for its clarity, scalability, and broad applicability. By utilizing cooperative awareness messages (CAMs) alongside a synthetic dataset, we map the topological relationships between vehicles. Our approach uses supervised Graph Neural Networks (GNNs) — specifically Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Graph Sample and Aggregate (GraphSAGE) — to accurately detect abnormal driving behaviors. The experimental results demonstrate that GraphSAGE outperforms GCN and GAT, achieving near-optimal performance with high recall, precision, and F1 scores in identifying these behaviors.