Enhancing Road Safety and Sustainability: A Multi-Scale Temporal Model for Vehicle Trajectory Anomaly Detection in Road Network Interactions

Effective anomaly detection in vehicle trajectories is crucial for developing sustainable and safe urban transportation systems. However, current research faces three main challenges including scarce anomaly data, inadequate spatial feature extraction in complex road networks, and limited capability in identifying complex behaviors. To address these issues, this paper proposes a Multi-scale Temporal and Road Network Interaction Anomaly Detection model (MTRI). Our framework leverages a Contrastive Learning-based Conditional Diffusion Model (CL-CD) to generate synthetic anomalous trajectories across diverse scenarios. It then employs an Urban road Network Interaction Modeling model (UNIM) to capture the profound interactions between trajectories and the road network. Finally, a Long-Short Temporal Anomaly Detection model (LSTAD) is designed to learn multi-scale temporal features for detecting sophisticated anomalies. Extensive experiments on real-world datasets from various urban scenarios demonstrate the superiority of our approach, which achieves high accuracy and adaptability (AUC-ROC > 0.85). This work contributes to sustainable urban mobility by providing a reliable solution for enhancing road safety through proactive anomaly detection.