Industrial safety has become a topic of great concern, especially in the oil and gas industry, where employee health and safety are of paramount importance. The correct wearing of work clothes and helmets is vital to employee safety, but there are still employees who do not comply. To solve this problem, many companies have adopted manual supervision and strengthened safety education, but the results are not ideal. Therefore, researchers developed an automatic helmet detection system. The development of target detection algorithms has gone through the era of manual feature construction and deep learning, in which single-stage target detection algorithms perform well in terms of speed and performance. However, existing methods are difficult to balance between detection efficiency and accuracy. This research aims to develop an all-weather real-time safety helmet and work clothes monitoring system, using automatic data collection methods to complete the collection of on-site data, saving a lot of manual screening time. The produced data set contains scenes in various time periods, and then the YOLOv7 network is used to train the data set, and Tensorrt is used to optimize the model. Experimental results show that the YOLOv7-based safety helmet and work clothes detection method MAP@O.5:0.95 has an accuracy of 80.42% and a YOLOv7 + Tensorrt detection speed of 90.39 FPS. YOLOv3 and YOLOv5 were used as comparative experiments to prove that our proposed helmet and work clothes detection method can meet the requirements of real-time detection environment while also having high detection accuracy.