Intelligent detection of safety helmets and reflective vests based on deep learning

Wearing a safety helmet and a reflective vest is essential for ensuring worker safety. While YOLO-based object detectors have demonstrated significant accuracy in detecting dress code violations, they often struggle with detecting small targets and maintaining a global focus. To address these challenges, we propose MSCG-YOLO, a novel algorithm based on YOLO networks for worker detection. Our approach integrates multi-head self-attention (MHSA) into the backbone network and neck connections, enhancing the model’s global field of view and its ability to detect small and obscured targets. To further improve small target detection, we designed a new neck structure called consolidative informative systematic neck (CISNeck), which includes additional layers and an enhanced detection head. We also developed the superficial feature fusion module (SFFM) to optimize the high-resolution features of the fourth detection head. Generalized intersection over union (GIoU) was used as the loss function. Experimental results on custom datasets show that MSCG-YOLO outperforms existing methods, achieving AP and AP50 values of 52% and 91.6% on the validation set, and 53.6% and 91% on the test set. Compared to YOLOv8n, MSCG-YOLO improves AP50 scores by 3.4% on the validation set and 2.7% on the test set. In conclusion, this study effectively addresses the practical needs of dress code detection in construction scenarios.