Development of Tool for Analysis of Swimming Using Pose Estimation Algorithm

Itay Coifman, May Hakim, Gera Weiss, Raziel Riemer

Abstract: Biomechanical analysis of swimming is notoriously difficult due to the aquatic environment. This study introduces an end-to-end system for analyzing swimming biomechanics using a single, uncalibrated underwater camera. We trained a YOLO-V8 Pose estimation model on manually labeled underwater videos of swimmers. To validate the system, we extracted key biomechanical features—including stroke rate, elbow angle at the push phase, and shoulder rotation—from a test set of transverse and sagittal videos and compared them to manual annotations. The results demonstrate high accuracy for temporal features like stroke rate (e.g., 0.19% MAPE) and promising results for 2D angle measurements (7-9 degree RMSE), despite challenges from 2D projection and video quality. This work represents a novel application of monocular computer vision for swimming analysis, showing that essential biomechanical data can be accurately extracted from simple video, paving the way for more accessible and larger-scale swimming studies.