SUMO-enhanced traffic anomaly detection: enhancing spatiotemporal transferability through simulation-based data augmentation

Traffic anomaly detection is crucial for road safety and traffic efficiency, but the scarcity of accident data poses challenges for model training. Although current research has utilized deep neural networks for data augmentation in small-sample scenarios, it has largely overlooked real-world physical constraints and lacks exploration into zero-shot scenarios, areas that still require further development. The paper proposes a SUMO-based simulation-driven data augmentation approach that enhances detection performance under limited real-world data conditions. Synthetic accident data are generated under five physical constraints and integrated with real traffic data to augment the training set. This study evaluates the method using three deep learning models (MLP, CNN, and LSTM) on PeMS datasets from two California highways (I80-E and I10-E). Experimental results demonstrate that the simulation-augmented training significantly improves model performance. In the spatial transfer experiment, the MLP model's accuracy improved from 75.53% to 76.58%, and AUC increased from 53.62% to 56.74%. These findings indicate that this approach effectively enhances both temporal and spatial transferability for traffic anomaly detection in small-sample scenarios.