Suppr超能文献

基于机器学习的羽毛球运动员动作运动识别模型分析

The analysis of motion recognition model for badminton player movements using machine learning.

作者信息

Zhu Xuanmin, Liu Lizhi, Huang Jingshuo, Chen Genyan, Ling Xi, Chen Yanshuo

机构信息

School of Physical Education, Fujian Polytechnic Normal University, Fuzhou, 350300, China.

Faculty of Education, Silpakorn University, Sanam Chandra Palace Campus, Amphoe Muang, Nakhon Pathom, 73000, Thailand.

出版信息

Sci Rep. 2025 May 30;15(1):19030. doi: 10.1038/s41598-025-02771-9.

Abstract

This study aims to comprehensively analyze and classify the badminton players' swing actions by combining the theoretical frameworks of quantum mechanics and machine learning. A badminton stroke recognition method based on Quantum Convolutional Neural Network (QCNN) is proposed. It is then compared with traditional Support Vector Machines (SVM) and Convolutional Neural Network (CNN). The comparison aims to assess the classification performance and robustness of each method. Firstly, this study collects the badminton players' stroke action data using high-frame-rate cameras and inertial sensors to record posture information during different strokes. OpenPose is used for human posture estimation, and combined with sensor data, spatiotemporal features during the stroke are extracted. Next, during the data preprocessing stage, Gaussian filtering is applied to remove noise, followed by normalization and feature selection to ensure the quality of the model input data. Then, SVM, CNN, and QCNN models are trained to classify different stroke actions. To evaluate model performance, precision, recall, and F1-score are selected as metrics. Experiments with varying noise levels (low, medium, and high noise) are designed to test the models' robustness. Finally, decision tree feature importance analysis is conducted to assess the contribution of different features to stroke action classification. Experimental results show that QCNN outperforms all other models in all classification tasks, with an F1-score of 0.860 for backhand intercept, significantly better than CNN (0.792) and SVM (0.753). In robustness tests under low, medium, and high noise environments, the classification precision of QCNN is 0.95, 0.92, and 0.89, respectively. This clearly surpasses both CNN and SVM. The results indicate that QCNN has strong adaptability to noisy data. Further feature analysis reveals that arm angle, twist angle, and step position are key factors affecting classification accuracy, with the highest contribution in the QCNN model. This study validates the superiority of QCNN in badminton action recognition and provides reliable methodological support for subsequent sports technique analysis.

摘要

本研究旨在结合量子力学和机器学习的理论框架,对羽毛球运动员的挥拍动作进行全面分析和分类。提出了一种基于量子卷积神经网络(QCNN)的羽毛球击球识别方法。然后将其与传统的支持向量机(SVM)和卷积神经网络(CNN)进行比较。该比较旨在评估每种方法的分类性能和鲁棒性。首先,本研究使用高帧率相机和惯性传感器收集羽毛球运动员的击球动作数据,以记录不同击球过程中的姿势信息。OpenPose用于人体姿势估计,并结合传感器数据,提取击球过程中的时空特征。接下来,在数据预处理阶段,应用高斯滤波去除噪声,然后进行归一化和特征选择,以确保模型输入数据的质量。然后,训练SVM、CNN和QCNN模型对不同的击球动作进行分类。为了评估模型性能,选择精度、召回率和F1分数作为指标。设计了不同噪声水平(低、中、高噪声)的实验来测试模型的鲁棒性。最后,进行决策树特征重要性分析,以评估不同特征对击球动作分类的贡献。实验结果表明,在所有分类任务中,QCNN均优于所有其他模型,反手截击的F1分数为0.860,明显优于CNN(0.792)和SVM(0.753)。在低、中、高噪声环境下的鲁棒性测试中,QCNN的分类精度分别为0.95、0.92和0.89。这明显超过了CNN和SVM。结果表明,QCNN对噪声数据具有很强的适应性。进一步的特征分析表明,手臂角度、扭转角度和步位是影响分类精度的关键因素,在QCNN模型中的贡献最大。本研究验证了QCNN在羽毛球动作识别中的优越性,并为后续运动技术分析提供了可靠的方法支持。

https://cdn.ncbi.nlm.nih.gov/pmc/blobs/f86f/12125242/c734b6b45480/41598_2025_2771_Fig1_HTML.jpg

文献AI研究员

20分钟写一篇综述,助力文献阅读效率提升50倍。

立即体验

用中文搜PubMed

大模型驱动的PubMed中文搜索引擎

马上搜索

文档翻译

学术文献翻译模型,支持多种主流文档格式。

立即体验