Machine learning for human pose assessment and correction in sports science

Human pose understanding has become a cornerstone for objective analysis and personalized feedback in sports science. Recent advances in machine learning (ML)—particularly deep learning for 2D/3D pose estimation, temporal modeling, and decision-making—enable accurate, non-invasive assessment of technique and real-time corrective guidance. This paper presents a consolidated treatment of ML for pose assessment and correction in athletic contexts. We formalize the pose quality estimation problem, summarize state-of-the-art approaches, define practical metrics, and propose a modular pipeline that combines efficient pose estimation, biomechanical feature extraction, and a cue policy for human-centered feedback. We discuss datasets, evaluation protocols, deployment on edge devices, privacy-preserving learning, and open challenges including domain shift, fairness, uncertainty, and human factors. Case studies in running gait, barbell back squat, and tennis serve illustrate end-to-end application from sensing to feedback.