Graph Attention Network Predicts Freeway Traffic Risk.

Assessing freeway traffic risk at a frame level is crucial for real-time safety monitoring, and this paper frames it as a multi-agent scene graph-level binary classification problem. The researchers construct a relation-aware graph for each video or trajectory frame, where vehicles are nodes and interactions (same-lane longitudinal, adjacent-lane lateral) are edges, enriched with physics-informed features. Building on a structured benchmarking suite, they propose HIA-GAT, a novel dual-stream heterogeneous graph attention network. HIA-GAT processes longitudinal and lateral interactions through dedicated attention pathways, fusing them via a conflict-type-aware gating mechanism supervised by event-level conflict attribution. Experiments conducted on the NGSIM I-80 and US-101 freeway datasets, across various conflict severity thresholds, demonstrate that HIA-GAT achieves the best average risk-ranking performance. Its gains are particularly significant in scenarios involving lane-change conflicts, where relational structure is essential. Beyond accuracy, the learned gate provides interpretable per-vehicle attribution of the dominant conflict type, offering actionable insights for real-time freeway safety management. The study highlights that graph structure is vital for modeling lateral conflict risk, while longitudinal risk can often be captured by simpler non-relational aggregations.