Abstract: Human unsafe behavior is one of the main causes of coal mine accidents.With the maturity of artificial intelligence technology represented by deep learning, the YOLOv5s object detection algorithm can also be applied to the intelligent recognition of unsafe behaviors in coal mines, but due to the complex underground environment of coal mines, the detection effect of YOLOv5s model is affected.We improve the YOLOv5s model, and boost the detection of small targets(such as hard hats) by adding a small target detection layer.Then, the DIoU loss function was used to replace CIoU to improve the accuracy of the object detection model.Finally, the CUMT-HelmeT dataset is trained, and the training results show that the mAP@0.5 of the improved algorithm reaches 90.4%,which is 1.8% higher than the original 88.6%.The improved detection algorithm lays a theoretical foundation for the intelligent identification of miners' unsafe behaviors.