HERMES graph AI predicts traffic conflicts with 95.7% detection rate
New heterogeneous graph model outperforms Transformers in intersection crash prediction...
Traditional surrogate safety measures (SSMs) treat traffic interactions as pairwise or flatten multi-agent scenes, losing structural and temporal nuance. To address this, Islam and Das introduce HERMES (Heterogeneous Edge-Relational Multi-Head Embedded SSM Attention), which formulates traffic conflict assessment as temporal heterogeneous scene-graph classification. Vehicles and pedestrians are represented as heterogeneous nodes with relation-specific edges capturing kinematic and safety descriptors. The model jointly learns relation-specific attention, dynamic node-edge updates, safety-aware graph pooling, and temporal sequences to estimate scene-level conflict probability.
HERMES was evaluated on 109,028 trajectory sequences from a signalized intersection and tested on an independent dataset. It achieved an AUC-ROC of 0.9898 ± 0.0013, AUC-PR of 0.9412 ± 0.0067, and F1 score of 0.8449 ± 0.0103. At a 5% false-alarm rate, it detected 95.7% of conflict sequences, outperforming the strongest Transformer baseline and XGBoost. In zero-shot external evaluation, it achieved AUC-ROC of 0.9752. Joint source-target training further improved performance with limited target-site data. These results demonstrate that preserving heterogeneous interaction topology and safety-informed edge semantics significantly improves scene-level conflict classification, enabling transferable roadside safety monitoring at signalized intersections.
- HERMES uses heterogeneous graph neural network with relation-specific edges for vehicle-vehicle, vehicle-pedestrian, and pedestrian-pedestrian interactions
- Achieves AUC-ROC 0.9898 and 95.7% conflict detection at 5% false-alarm rate, outperforming Transformer and XGBoost baselines
- Zero-shot external evaluation shows AUC-ROC 0.9752, demonstrating transferability to new intersections without retraining
Why It Matters
HERMES enables proactive, real-time traffic conflict prediction at intersections, improving safety for autonomous vehicles and smart city infrastructure.