Human pose estimation in powerlifting

Abstract

Amidst the rising popularity of powerlifting, the need for advanced techniques to maintain proper form and prevent injuries has become increasingly crucial. This study explores the burgeoning field of powerlifting, leveraging AI‐driven solutions to enhance movement analysis and support lifters in optimising their technique. Our approach focused on curating a dataset to train a powerlifting‐specific Human Pose Estimation (HPE) model, ensuring diverse representations of squats, bench presses, and deadlifts. The model was trained using the You Only Look Once (YOLO) framework, leveraging manually labelled keypoints as ground truth data for both training and evaluation. The model demonstrated strong performance across multiple evaluation metrics, confirming its effectiveness in powerlifting analysis. It achieved a Percentage of Correct Parts (PCP) of 89.79%, a Percentage of Detected Joints (PDJ) of 97.16%, and a Percentage of Correct Keypoints (PCK) of 91.31%, highlighting its high precision in keypoint detection. Additionally, the Mean Per Joint Position Error (MPJPE) of 14.25 pixels and Mean Absolute Joint Angle Error (MAJAE) of 6.16 degrees reflect its accuracy in localising joints and estimating movement angles. To ensure AI‐driven analysis translates effectively into practical use, a UI was developed, allowing users to upload media, visualise detected keypoints, and receive structured feedback. The trained YOLO model was integrated into a process that includes additional calculations for perspective classification and form analysis, achieving 92.76% and 90.06% accuracy, respectively. These insights were processed using the ChatGPT API to generate contextually relevant and actionable feedback. Designed as a proof of concept, the UI demonstrates the potential of AI‐powered form feedback in real‐world applications. Extensive testing validated its usability and robustness, ensuring smooth functionality to support lifters in refining their technique. The system establishes a strong foundation for advancing AI‐driven sports analysis, enabling refinements in real‐time assessment, dataset expansion, and model accuracy. By bridging the gap between AI technology and practical athletics, this research highlights its transformative potential in performance analysis. Aligned with global health and fitness goals, it underscores AI’s role in fostering inclusivity, promoting safer training methodologies, and setting new standards in the application of AI for enhancing sports performance and safety.