As the most basic personal protective equipment, helmets are of great significance to the safety for workers. However, some workers lack safety awareness and often do not wear safety helmets. This study focuses on the detection of safety helmet in complex background. You Only Look Once (YOLO) is a state-of-the-art, real-time object detection algorithm, we propose to apply the YOLO detector for safety helmets detection, which achieves high accuracy. For the single-type detection problem without wearing a helmet, the classifier is modified and the output is modified to a tensor of 18 dimensions. We train YOLOv3 for safety helmets detection on the 2010 datasets based on the pre-training model in ImageNet. Then we optimize the model according to the loss function and IOU curve. The experimental results show that the safety helmet detector gets 98.7% accuracy and 35 fps on the 2000 detection test sets without GPU, which meets the real-time detection requirements. The effectiveness of the YOLOv3 safety helmet detection method is verified.