As a vital part of traditional Chinese medicine, acupuncture has extensive application value all over the world. However, the reliance on practitioners’ experience for acupoint localization in traditional acupuncture methods leads to a lack of standardization, restricting its reproducibility and broader adoption. Acupuncture robots, as a kind of intelligent medical devices, offer new opportunities for standardizing and promoting acupuncture techniques. This paper introduces an improved YOLOv8-Pose model, YOLO-PointMap, designed to address challenges in dense acupoint distribution and weak feature recognition. By incorporating dynamic convolution to optimize the C2f module and introducing a channel-attention-based feature fusion module, the model achieves significant advancements in multi-scale feature extraction and integration. Experimental results show that the end point error (EPE), percentage of correct keypoints (PCK) and mAP50-95(Pose) indexes of YOLO-PointMap on the test set are superior to the existing models, with the values reaching 3.27, 1.00 and 84.90% respectively, especially in dense key point identification and weak feature region localization. It provides strong support for the development of acupuncture robot technology, and shows the potential application value in the fields of virtual reality and intelligent interaction.