基于LSA-HRnet网络的人体姿态估计方法在太极拳运动中的应用  

Application of human pose estimation method in Tai Chi exercise based on LSA-HRnet network

作　　者：徐广 吴星辰 XU Guang;WU Xingchen(Sports Department,Shenyang Aerospace University,Shenyang 110136,China;Jiushao Institute of AI Algorithm,Jihua Laboratory,Foshan 528200,China)

机构地区：[1]沈阳航空航天大学体育部综合教研室,沈阳110136 [2]季华实验室九韶人工智能算法研究院,佛山528200

出　　处：《中南民族大学学报（自然科学版）》2023年第6期839-845,共7页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：沈阳航空航天大学教改资助项目(JG2022120)。

摘　　要：由于专业运动知识的匮乏、动作评估准确度低等问题,导致人们对运动训练的积极性不高,运动水平提升慢以及身体出现不同程度的损伤等,针对以上问题,提出一种基于注意力机制的轻量级采样模块的HRnet(Lightweight Sampling Attention block HRnet,LSA-HRnet)人体姿态估计模型对运动过程中的动作进行分析和评估,提出的算法采用轻量级采样块网络和融合注意力机制实现模型的轻量化以及模型性能的提升.相比原始HRnet模型及其他的优秀人体姿态估计SCnet模型、轻量级Lite-HRnet,在自制太极拳的实验结果表明提出的模型能够有效的提高预测精度和降低参数量.基于提出方法能丰富混合现实技术在运动领域的发展的技术理论,改进现存运动问题、激发练习兴趣和提升体质健康.Insufficient knowledge of professional sports and low accuracy in motion assessment have led to low motivation in sports training,slow improvement in athletic performance,and various degrees of physical injuries.To address these challenges,a lightweight human pose estimation model,namely Lightweight Sampling Attention block HRnet(LSAHRnet)is proposed,based on an attention mechanism,for analyzing and evaluating during sports activities.The proposed algorithm combines lightweight sampling block networks and fusion attention mechanisms to achieve model light weighting and performance enhancement.Comparative experiments with the original HRnet model,as well as other notable human pose estimation models such as SCnet and Lite-HRnet,are conducted using a self-designed Tai Chi dataset.The experimental results demonstrate that the proposed model effectively improves prediction accuracy while reducing the parameter count.Moreover,based on the proposed approach,the research enriches the technical theory for the development of mixed reality technology in the field of sports,improves existing sports-related issues,stimulates exercise interest,and enhances physical fitness and health.

关 键 词：HRnet模型 人体姿态估计 混合现实 太极拳运动 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]