Model-Agnostic Meta-Learning-Based Real-Time Traffic Conflict Prediction With Limited Sample at Heterogeneous Signalized Intersections

Machine learning-based real-time traffic conflict prediction has been a crucial part for proactive traffic safety management in Intelligent Transportation System. However, most existing models struggle with limited-sample data and poor generalization across different locations. This study proposes a Model-Agnostic Meta-Learning (MAML)-driven deep learning framework that integrates Transformer and Long Short-Term Memory (LSTM) architectures to address the challenges of real-time conflict prediction at signal cycle level for intersections. The MAML framework enables rapid adaptation to new intersections by learning a universal initialization from meta-tasks, while the Transformer-LSTM model captures complex relationships in multivariate traffic parameters and temporal variations in traffic flow characteristics. Standalone LSTM and Transformer models are employed as baselines and to validate the model-agnostic nature of MAML. Video recordings from four signalized intersections were used to extract key traffic features and lane configuration was incorporated as additional input features. Severe rear-end conflict events were labeled based on the 85th percentile threshold of the Modified Time to Collision (MTTC). Extensive comparative experiments show that the MAML-driven models significantly outperform their non-MAML counterparts. Among them, the MAML-driven Transformer-LSTM model significantly outperforms both MAML-driven standalone LSTM and Transformer models. The results underscore MAML’s advantage in implicitly capturing intersection heterogeneity and its effectiveness in handling traffic conflict prediction under limited-sample scenarios. Additionally, sensitivity analysis revealed that the model showed low sensitivity to the sliding time window and the percentile defining severe traffic conflict thresholds, while transferability analysis confirmed its ability to generalize across four intersections and external data from different urban region.