Monitoring unsafe behaviors of miners in complex environments

Unsafe behaviors of miners are a primary cause of coal mine accidents, and intelligent video monitoring is an important means of preventing such accidents. To improve unsafe-behavior recognition in complex underground environments with low illumination, motion blur, occlusion and dense personnel, this study proposes an improved YOLOv12n-based detection model. Video data from coal mining faces, transportation roadways and other key areas were filtered and annotated to construct a two-class dataset containing the unsafe-behavior categories of no-helmet and no-mask. DnCNN preprocessing was used to suppress image noise and enhance visual saliency. On the YOLOv12n baseline, three modules were integrated for different purposes: a Discriminative Frequency Domain-based Feed Forward Network (DFFN) was introduced into the neck to strengthen frequency-domain feature representation and blurred-target perception; Dynamic Tanh (DyT) was used to replace normalization operations in selected backbone/neck blocks to improve training stability and multi-scale feature adaptation; and the Mona module was embedded in the feature-fusion stage to enhance multi-scale contextual aggregation for occluded targets. Ablation and comparative experiments were conducted using precision, recall, F1-score, mAP50 and mAP50-95 as evaluation metrics. The proposed YOLOv12n-DFFN-DyT-Mona model achieved 92.16% precision, 90.46% recall, 91.30% F1-score, 94.27% mAP50 and 61.30% mAP50-95, outperforming the YOLOv12n baseline by 2.36, 3.45, 2.92, 2.35 and 2.27 percentage points, respectively. Representative visual comparisons further indicate that the proposed model reduces missed detections under occlusion, crowding and poor illumination.