HIA-GAT Predicts Freeway Traffic Conflict Risk with High Accuracy

Researchers have developed HIA-GAT, a novel Heterogeneous Interaction-Aware Graph Attention Network, designed to predict traffic conflict risk on freeways at a frame-by-frame level. This model frames the problem as a multi-agent scene graph-level binary classification, where each video or trajectory frame is labeled as risky if any conflict, based on Time-to-Collision (TTC) or Post-Encroachment Time (PET) thresholds, is detected. HIA-GAT constructs a relation-aware graph for each frame, treating vehicles as nodes and defining two types of interaction edges: longitudinal (same-lane) and lateral (adjacent-lane). These edges are enriched with physics-informed features tailored to rear-end and lane-change conflict mechanisms. The network employs a dual-stream architecture, processing longitudinal and lateral interactions through dedicated attention pathways. These streams are then fused using a conflict-type-aware gating mechanism, supervised by event-level conflict attribution. Evaluated on the NGSIM I-80 and US-101 freeway datasets across various TTC and PET threshold configurations, HIA-GAT consistently achieved the best average risk-ranking performance, with AUC scores of 0.835 on I-80 and 0.867 on US-101. The model showed particular strength in PET-only (lane-change) settings, where understanding relational structure is crucial. Beyond its accuracy, the learned gate provides interpretable per-vehicle attribution of the dominant conflict type, offering actionable insights for real-time freeway safety monitoring.