Abstract: In order to effectively detect and identify the dangerous behaviors of underground workers in coal mines and prevent the occurrence of safety accidents, a detection algorithm for dangerous actions of miners based on attention mechanism and deep learning is proposed for the problems of complex background and large scale changes in coal mines. On the basis of the YOLOv3 model, a lightweight feature extraction network with stronger feature extraction capability and smaller volume is designed; in view of the poor detection performance of the original YOLOv3 algorithm on small targets, an attention-based approach is proposed. The feature fusion module of the mechanism optimizes the missed detection and false detection of small targets. In order to evaluate the performance of the model, 10 000 underground coal mine pictures were collected for training and testing. The mAP of the algorithm proposed in this paper is 83.1%, which is 6.6 times higher than the current commonly used target detection algorithms. In addition, the algorithm test speed is 769 fps, which is faster than other light weight object detection algorithms. The experimental results prove that the dangerous behavior detection algorithm proposed in this paper can be applied to the actual production environment.