Safety Helmet and Mask Detection at Construction Site Based on Deep Learning

To reduce safety hazards and effectively detect the wearing of helmets and masks by construction workers in complex scenarios, an improved YOLOv4-based detection algorithm called YOLOv4-HelMask was proposed in this paper. It includes dataset construction and data augmentation, k-means algorithm-based check frame selection, PANet-based structure separation category prediction, and location prediction to enhance feature extraction. The convergence of the model was optimized using a cosine annealing controller, and the generalization of the model was enhanced using a label smoothing technique. Experimental results demonstrate that the proposed YOLOv4-HelMask algorithm achieves a mAP value of 95.51% on the self-built dataset, which is 21.99%, 5.89%, 2.42%, and 5.07% higher than the current representative Faster R-CNN, SSD, YOLOv5, and YOLOv4 algorithms, respectively. The proposed algorithm can meet the needs of real-time detection.