Abstract: Detection of dangerous behaviors of personnel in underground mines is a critical aspect of coal mine safety and risk prevention. Existing object detection technologies face challenges when applied to underground personnel behavior detection, as complex working conditions, equipment occlusion, dense targets, and dust interference often lead to inaccurate feature extraction and undefined behavior classifications. To address these issues, a PMR-YOLO model was constructed based on the YOLOv8-pose architecture. The standard convolution was replaced with a hybrid DCNv4-PConv module, which combines the DCNv4 network and the deformable capability of the PConv module. In addition, a Mixed Local Channel Attention (MLCA) module was integrated into the structure, and the detection head was replaced with a Receptive-Field Attention Convolution (RFAConv) module. These modifications aimed to improve the accuracy and speed of human keypoint detection in underground surveillance images. On this basis, a personnel behavior recognition algorithm was designed to classify underground behaviors into nine categories. Human skeletal structures were generated from the keypoints detected by the YOLOv8-pose model, and behavior types were identified accordingly. The DsLMF+ dataset was used to conduct ablation experiments, comparative experiments, and behavior recognition experiments. The results showed that incorporating the DCNv4-PConv hybrid module, the MLCA module, and the RFAConv module significantly improved the precision, recall, and Mean Average Precision (mAP) of the YOLOv8-pose model. The PMR-YOLO model achieved a precision of 0.893, a recall of 0.841, and an mAP of 0.852 in keypoint detection, representing improvements of 6.9%, 14.4%, and 10.5%, respectively, over the original YOLOv8-pose model. The detection method based on the PMR-YOLO model can effectively identify nine types of underground personnel behaviors, and all recognition accuracies exceed 96%.