Abstract: Unsafe behavior management is an important part of coal mine safety management. Integrating modern tech⁃ nologies to realize the intelligent recognition of unsafe behavior of coal mine employees is of great significance to the intelligent construction of coal mine safety management. In order to realize the intelligent recognition of underground employees' unsafe behaviors，the unsafe behaviors are divided into static unsafe behaviors，dynamic unsafe behaviors and interactive unsafe behaviors. Then，an unsafe behavior recognition system based on computer vision is designed. The system includes six subsystems:video acquisition，data warehouse，video processing，model management，rule rea⁃ soning and early warning system，which correspond to the six functions that include the aggregation and management of underground monitoring video，data storage，static and dynamic unsafe behavior recognition，model management，inter⁃ active unsafe behavior identification and early warning trigger，and early warning management. The whole system uses Open Pose neural network to identify the key points of human and human body，uses YoloV3 neural network to identify equipment and environment，uses MobileNetV3 neural network to identify static unsafe behavior，uses ST⁃GCN neu⁃ ral network to identify dynamic unsafe behavior and uses rule reasoning method to identify interactive unsafe behavior. The neural network model is written using Python + pytorch， and uses public data set and special data set for pre⁃training and formal training. The accuracy of static unsafe behavior，dynamic unsafe behavior and interactive unsafe behavior recognition of the whole system is up to 89.4%，90.7% and 75.6% respectively. Finally，based on ku⁃ bernetes + docker container，the deployment of the identification system in the actual production environment is real⁃ ized. The whole research integrates a variety of methods to the identification of multiple unsafe behaviors，and has been applied in the actual production environment.