Lane-Changing Safety Assessment Using Graph Attention Networks for Autonomous Driving

Lane-changing is a safety-sensitive maneuver in driving, directly influencing both traffic efficiency and collision risk. Existing rule-based and learning-based methods struggle to capture complex inter-vehicle dependencies, motivating our graph-based approach. In this paper, we propose a graph attention network-based framework for lane-changing safety assessment using the NGSIM dataset. We first detect lane-changing events and extract decision moments through a refined filtering and event identification method. We then construct traffic scene graphs representing spatiotemporal relationships between the ego vehicle and surrounding vehicles. A multi-head GAT architecture is designed to adaptively weight the influence of neighbors, enabling both discrete safety classification (safe, risky, dangerous) and continuous risk scoring. Experiments show that our method achieves superior accuracy and F1-score compared with baseline models such as LSTM and GCN. This shows that explicitly modeling inter-vehicle interactions with attention mechanisms significantly enhances safety evaluation for advanced driver assistance systems and autonomous driving.