Validating a generative AI-assisted pose refinement framework for kinematic analysis in elite taekwondo athletes.

This study evaluated a Generative AI-Assisted Pose Refinement Framework integrating Large Language Models (LLM) and Retrieval-Augmented Generation (RAG) to enhance markerless motion capture accuracy in elite Taekwondo kyorugi athletes. Eight elite athletes (age: 21.1 ± 0.9 yrs) underwent a 12-week training macrocycle. Kicking kinematics were recorded using MediaPipe Pose, with the LLM-RAG system developed to detect and correct anatomical outliers based on biomechanical constraints. Results indicated that the framework successfully identified 10.37% of keypoints as outliers, improving detection accuracy by 5.69% to 13.24% compared to raw MediaPipe estimates. Furthermore, the refined data effectively captured individualised performance adaptations following the intervention: five athletes demonstrated significant improvements in stability (with the coefficient of variation (CV) for speed decreasing by up to 39.66%), one maintained performance, and two exhibited increased variability, validating the system's sensitivity to longitudinal training effects. This approach significantly enhances the reliability of markerless tracking for high-velocity movements, providing a viable, field-based alternative for continuous biomechanical monitoring in applied sports settings.