In construction, it is important to check whether workers wear safety helmets in real time. We proposed using an unmanned aerial vehicle (UAV) to monitor construction workers in real time. As the small target of aerial photography poses challenges to safety-helmet-wearing detection, we proposed an improved YOLOv4 model to detect the helmet-wearing condition in aerial photography: (1) By increasing the dimension of the effective feature layer of the backbone network, the model's receptive field is reduced, and the utilization rate of fine-grained features is improved. (2) By introducing the cross stage partial (CSP) structure into path aggregation network (PANet), the calculation amount of the model is reduced, and the aggregation efficiency of effective features at different scales is improved. (3) The complexity of the YOLOv4 model is reduced by introducing group convolution and the pruning PANet multi-scale detection mode for de-redundancy. Experimental results show that the improved YOLOv4 model achieved the highest performance in the UAV helmet detection task, that the mean average precision (mAP) increased from 83.67% of the original YOLOv4 model to 91.03%, and that the parameter amount of the model is reduced by 24.7%. The results prove that the improved YOLOv4 model can effectively respond to the requirements of real-time detection of helmet wearing by UAV aerial photography.

No abstract