Enhanced detection and classification of unsafe behaviors in hydropower stations via computer vision techniques

Unsafe behaviours among hydropower station personnel contribute to accidents and are closely associated with ergonomic risks. To address inefficient manual monitoring and limited consideration of behavioural risks in safety management, this study proposes an improved YOLOv11n-based method for unsafe behaviour detection and develops an ergonomic risk assessment framework using detection results. The method incorporates PConv, an attention mechanism, and Focal Loss into YOLOv11n and integrates YOLOv11n-Pose and DCNN-SE modules to identify unsafe behaviours involving personal protection and actions. Experimental results demonstrate improvements over the comparison model in precision, recall, mean average precision, and inference speed, with robustness to complex backgrounds, small targets, and partial occlusion.