Cardiopulmonary resuscitation (CPR) is critical for improving the survival rate of patients with cardiac arrest. Traditional CPR training and assessment mainly rely on manual observation, which suffers from strong subjectivity and a lack of quantitative metrics. Moreover, existing methods show insufficient accuracy in handling the severe limb occlusions and fine-grained motion capture that are characteristic of CPR. To address these challenges, this paper proposes a CPR pose estimation algorithm based on an improved YOLOv13n-Pose model. The proposed method is validated on a self-constructed CPR pose estimation dataset covering multiple subjects and scenarios.First, to tackle the severe self-occlusion formed by the rescuer’s arms and torso, a Bi-Level routing attention (BRA) mechanism is introduced, which enhances the network’s global feature extraction capability in complex scenes through dynamic sparse perception. Second, a hybrid enhanced upsample (HEUpsample) module is designed to replace the neck upsampling structure. By incorporating spatial offset and channel shuffle strategies, the module reduces detail loss during feature map upsampling and improves keypoint localization accuracy. Experimental results demonstrate that the improved model achieves mAP scores of 95.8% and 88.5% for CPR action recognition and pose estimation on the CPR dataset, representing improvements of 2.3% and 3.6% over the baseline model, respectively. Compared with mainstream methods, the proposed approach maintains a high inference speed while reducing computational cost, and it also demonstrates strong generalization ability on the public OCHuman dataset.