Component Replacement Study of 3D Human Pose Estimation Models in Real-World Complex Sports Scenarios: Focusing on Head Impact Events

Yuchen Shi, Nobutake Ozeki, Ryu Yoshida, Motoki Inaji, Kazuyoshi Yagishita, Yusuke Miyazaki

2024

Abstract

Human pose estimation in 3D is crucial in complex sports scenarios, particularly for athlete head impact events. We investigated the effect of different 2D pose estimation methods on the performance of 3D pose estimation models in complex sports environments. We used a transformer-based 3D human pose estimation model as a base framework, creating multiple variants by replacing the 2D pose estimator. These variants were evaluated using real sports game videos. Four 2D pose estimators were employed: Simple Baseline, High-Resolution Network (HRNet), Multi-stage Pose Network (MSPN), and Residual Steps Network (RSN). Performance was assessed using Mean per Joint Positional Error (MPJPE), Procrustes analysis MPJPE (P-MPJPE), and Mean per Joint Velocity Error (MPJVE) metrics. The results showed that MSPN performed the best in terms of position accuracy and motion velocity consistency (MPJPE, P-MPJPE and MPJVE). RSN presented promising absolute position accuracy (MPJPE) but showed limitations in the overall pose configuration (P-MPJPE). Simple Baseline and HRNet proved to be inadequate for complex sports scenarios. These findings indicate that different model architectures have different advantages in 3D human pose estimation in complex sports scenarios. This study provides insights for improving 3D pose estimation models in challenging real-world sports applications, contributing to the better understanding and prevention of sports-related head injuries.

Paper Citation

in Harvard Style

Shi Y., Ozeki N., Yoshida R., Inaji M., Yagishita K. and Miyazaki Y. (2024). Component Replacement Study of 3D Human Pose Estimation Models in Real-World Complex Sports Scenarios: Focusing on Head Impact Events. In Proceedings of the 12th International Conference on Sport Sciences Research and Technology Support - Volume 1: icSPORTS; ISBN 978-989-758-719-1, SciTePress, pages 72-82. DOI: 10.5220/0013005000003828

in Bibtex Style

@conference{icsports24,
author={Yuchen Shi and Nobutake Ozeki and Ryu Yoshida and Motoki Inaji and Kazuyoshi Yagishita and Yusuke Miyazaki},
title={Component Replacement Study of 3D Human Pose Estimation Models in Real-World Complex Sports Scenarios: Focusing on Head Impact Events},
booktitle={Proceedings of the 12th International Conference on Sport Sciences Research and Technology Support - Volume 1: icSPORTS},
year={2024},
pages={72-82},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013005000003828},
isbn={978-989-758-719-1},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 12th International Conference on Sport Sciences Research and Technology Support - Volume 1: icSPORTS
TI - Component Replacement Study of 3D Human Pose Estimation Models in Real-World Complex Sports Scenarios: Focusing on Head Impact Events
SN - 978-989-758-719-1
AU - Shi Y.
AU - Ozeki N.
AU - Yoshida R.
AU - Inaji M.
AU - Yagishita K.
AU - Miyazaki Y.
PY - 2024
SP - 72
EP - 82
DO - 10.5220/0013005000003828
PB - SciTePress