Abstract: To address the issue of insufficient detection accuracy of lightweight pose estimation models for fine-grained keypoints such as the ankle joint, we propose an improved lightweight keypoint detection model, YOLOv8-FT. Built upon the YOLOv8n-pose baseline, the model incorporates three core improvements: introducing the FasterNet module based on partial convolution to reduce model complexity; embedding the efficient multi-scale attention (EMA) mechanism to enhance the extraction of subtle features; designing the HRNet_FT multi-resolution feature fusion network to optimize the retention and fusion of ankle keypoint features. Evaluated on a self-constructed ankle keypoint detection dataset, YOLOv8-FT achieves an average precision (mAP50-95) of 85.2%, an 11.6% improvement over the baseline, with an ankle joint angle measurement error of less than 3°. Experimental results demonstrate that the model significantly improves the detection accuracy of ankle keypoints while maintaining lightweight and real-time performance, enabling stable tracking in complex motion scenes. It provides a lightweight solution that balances accuracy and efficiency for mobile motion analysis and rehabilitation assessment.