A graph-based spatio-temporal framework for predicting safety-critical pedestrian-vehicle interactions at...

Pedestrian safety remains a critical global concern, especially in countries like India, where unsignalized crossings with limited traffic control contribute to high pedestrian fatality rates. This study proposes a novel graph-based framework for analyzing pedestrian-vehicle interactions, advancing beyond traditional indicator-based, trajectory prediction, and pairwise modeling approaches. Indicator-based methods (e.g., PET, TTC) are retrospective and fail to capture evolving dynamics. Trajectory prediction models suffer from error accumulation, while pairwise models like LSTM or Transformer architectures are limited to two-agent interactions, restricting scalability and scene comprehension. In contrast, the proposed framework constructs holistic, scene-level multi-relational graphs, representing pedestrians and vehicles as interconnected nodes, with explicit modeling of pedestrian-pedestrian, vehicle-vehicle, and pedestrian-vehicle interactions. Unlike existing approaches that treat each pedestrian-vehicle pair in isolation, the proposed method represents all road users present in the scene as nodes in a dynamic spatio-temporal graph, enabling the model to learn not only direct pedestrian-vehicle relationships but also indirect influences mediated by surrounding pedestrians and vehicles. This design eliminates the need to pre-select interaction pairs, simplifying real-world deployment and improving scalability in dense traffic scenarios. Disentangled Multi-Scale Aggregation (DMSA) captures group behavior by focusing on contextually relevant agents, while a temporal CNN backbone models both short- and long-range dependencies efficiently. Empirical evaluations demonstrate the superior performance of the proposed model, which achieved a test accuracy of 90.6%, F1-score of 0.906, precision of 0.927, recall of 0.886, specificity of 0.928, and an AUC of 0.950, outperforming widely used baselines from the literature, such as GRU, LSTM, and Transformer-MLP, that have been applied in pedestrian interaction modeling tasks. Ablation studies confirmed the importance of Multi-Relational Adjacency Matrices (MRAM) and DMSA in improving accuracy and reducing false positives. By modeling scene-level dynamics, the framework enables context-aware prediction of critical events, supporting proactive conflict warning systems.