Human pose estimation for automated biomechanical swimming analysis

Human pose estimation (HPE) has shown promise in biomechanical analysis across various sports, but its application to swimming has been limited due to the challenges of underwater video analysis. This study presents a novel system for analysing swimming biomechanics using single-camera videos and HPE. We developed specialised models for sagittal (side) and transverse (front) views, achieving average precision (AP) scores of 91.5% and 91.6%, respectively, thus significantly outperforming out-of-the-box HPE models. The system automatically extracts key biomechanical parameters, including stroke rate, kick rate, projected joint angles, and stroke phase timings. The model was tested on a dataset of 24 videos obtained from 12 different swimmers (1 video per view) demonstrating strong agreement with manual annotations, with mean absolute percentage errors (MAPE) of less than 1% for both stroke rate and kick rate. Phase detection accuracy (RMSE) ranged from 1.08 to 2.49 frames, depending on the stroke phase and camera view. The system enables automated analysis of swimming technique without requiring complex motion capture equipment or extensive human annotation, allowing for large-scale biomechanical swimming studies. The system's accuracy and accessibility make it a promising tool for both research and practical applications in swimming biomechanics.