DSR-YOLOv8

The power construction industry's growth demands efficient monitoring of high-risk worker behaviors, yet traditional methods are inefficient and existing models face false alarms in complex scenes. This study proposes DSR-YOLOv8, an improved YOLOv8 algorithm integrating three modules: (1) DSRAB using deep separable convolution and global pooling to enhance subtle action features and denoising; (2) SD_SPPF with multi-scale dilated kernels to expand the receptive field while reducing computational costs; (3) dynamic region-processing with partial convolutional heads to focus on critical areas and suppress interference. Evaluated on a self-built Dangerous Behavior Dataset (DBD) containing “helmet-wearing,” “no helmet,” and “smoking” scenarios, DSR-YOLOv8 achieved 91.2% accuracy (+3.5%) and 89.7% mAP (+3.6%) over baselines, demonstrating efficient hazardous behavior detection for enhanced safety in power construction.