基于人体姿态估计的羽毛球动作评估方法研究  

Study on badminton movement evaluation method based on human pose estimation

作　　者：赵小虎[1,2] 王轲鑫 孟献峰 石传寿[1,2] ZHAO Xiaohu;WANG Kexin;MENG Xianfeng;SHI Chuanshou(National and Local Joint Engineering Laboratory of Mine Internet Application Technology,China University of Mining and Technology,Xuzhou 221008,Jiangsu China;School of Information and Control Engineering,China University of Mining and Technology,Xuzhou 221008,Jiangsu China;School of Physical Education,China University of Mining and Technology,Xuzhou 221008,Jiangsu China)

机构地区：[1]中国矿业大学矿山互联网应用技术国家地方联合工程实验室,江苏徐州221008 [2]中国矿业大学信息与控制工程学院,江苏徐州221008 [3]中国矿业大学体育学院,江苏徐州221008

出　　处：《华中科技大学学报（自然科学版）》2024年第11期110-116,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：中央高校基本科研业务费专项资金资助项目(2020ZDPY0223)。

摘　　要：针对传统运动动作捕捉分析方法实时性不足、动作评估准确度低等问题,将人体姿态估计技术应用于羽毛球运动教学中,对羽毛球挥拍动作的识别与标准程度评估进行研究.首先,提出一种基于改进OpenPose的轻量级人体姿态估计模型,将VGG19特征提取网络替换为轻量型的MobileNet网络,并将模型内部的7×7卷积核结构改造为由一个1×1的卷积、一个3×3的深度可分离卷积与一个膨胀系数为2的空洞卷积组成的串联结构,实现了模型的轻量化;然后,针对羽毛球挥拍动作特点,提出14点人体姿态的稀疏表示模型对挥拍动作标准程度进行评估.实验结果表明:改进后的OpenPose模型在将性能提升3倍的同时,对人体手臂骨骼点的识别准确率提升了3.57%;14点人体姿态稀疏表示模型在保证精度的前提下,时效性提升了2.98倍.Aiming at the issues of insufficient real-time performance and low accuracy in traditional motion capture analysis methods,human pose estimation technology was applied to badminton sports instruction,the identification and standard evaluation of badminton swing movements were studied.First,an improved human pose estimation model based on OpenPose was proposed.T he VGG19 feature extraction network was replaced with a lightweight MobileNet network,and the 7×7 convolution kernel structure inside the model was restructured into a tandem structure composed of a 1×1 convolution,a 3×3 depthwise separable convolution,and a dilated convolution with a dilation rate of 2,achieving model lightweighting.Then,according to the characteristics of badminton swing actions,a sparse representation model of 14-point human pose was proposed to evaluate the standard degree of swing actions.Experimental results show that the improved OpenPose model increases processing speed by 3 times while simultaneously improving the accuracy of skeletal joint recognition for human arms by 3.57%,and the 14-point human pose sparse representation model achieves a 2.98-fold improvement in response time while maintaining precision.

关 键 词：人体姿态估计 羽毛球训练 人体姿态描述 姿态相似度计算 OpenPose模型 

分 类 号：TP391.6[自动化与计算机技术—计算机应用技术]