YOLOv11-Safe: An Attention-Enhanced and Explainable Framework for Campus Crisis Detection and Risk Level Prediction

Campus environments are increasingly exposed to sudden health incidents, violent conflicts, and abnormal behaviors, which challenge conventional surveillance systems due to their limited real-time performance and interpretability. This study proposes an enhanced YOLOv11-Safe framework that integrates an improved SimAM attention mechanism with a normalized Wasserstein distance (NWD) loss for multi-scenario crisis detection. A dedicated dataset of 2,500 images was constructed, covering seizures, fighting, fainting, vomiting, and wheelchair accidents. Experimental evaluation demonstrates that YOLOv11-Safe achieves a balanced trade-off between accuracy and efficiency while remaining lightweight for deployment. For risk-level prediction, a random forest model with SHAP interpretability analysis identified action frequency as the dominant feature, with event duration and crowd density providing additional contributions. Cross-validated PR and ROC–AUC curves further verified the robustness and generalization ability of the proposed framework. Overall, this work combines deep detection with interpretable risk analysis to provide a reliable basis for intelligent safety management in higher-education institutions.