Abstract: The unsafe behavior of underground operators is one of the main causes of mine accidents. The existing underground monitoring methods are still mainly manual monitoring,which can not quickly identify the unsafe behavior of operators, resulting in the difficulty of real-time early warning. A deep learning-based development scheme of unsafe behavior identification and early warning system for underground personnel is designed. Firstly,the data set based on underground environment is made,and then the YOLOv4 network model is used to identify miners and safety helmets to judge the wearing of safety helmets. Then,openpose algorithm and ST-GCN model are adopted to identify the miners′ behavior in the monitoring video. Finally,the system is used to automatically warn the unsafe behavior. Combining various deep learning and computer development techniques,a kind of identification and early warning system for unsafe behavior of underground personnel is developed,which provides a new idea for safety management of underground personnel,at the same time,it has reference significance for the construction of intelligent mine.