基于BlazePose的网球发球人体关节识别研究  

Research on Human Jjoint Recognition of Tennis Serve Based on BlazePose

作　　者：杨林峰 宗志伟[2] 

机构地区：[1]上海大学通信与信息工程学院,上海200444 [2]上海大学体育学院,上海200444

出　　处：《工业控制计算机》2023年第11期115-117,120,共4页Industrial Control Computer

摘　　要：网球发球时人体肢体动作过快,导致现有方法在网球发球姿态检测中存在关节识别错检率高的问题,由此提出了一种基于BlazePose的网球发球人体关节识别方法。该方法在BlazePose检测模型的基础上,设计了帧间姿态运动距离梯度鉴别算法,分为平缓运动、剧烈运动及自遮挡三种错误帧,分别利用光流肢节预测、高斯混合模型前景消除及三次样条插值,对错误帧进行对应修复。将该方法与BlazePose及lightweight-OpenPose进行对比实验,结果表明,该方法的整体识别准确率比BlazePose与lightweight-OpenPose分别高出3.1%与7.8%,并且对挥拍臂的腕关节和肘关节的检测有较大改善,执行效率与BlazePose相当,能辅助网球发球的日常训练。This paper proposes a method of tennis serving human joint recognition based on BlazePose.Based on the BlazePose detection model,this method designs a method to identify the distance gradient of pose motion between frames.It classifies three kinds of error frames:gentle motion,violent motion and self-occlusion.It uses optical flow limb prediction,Gaussian mixture model foreground elimination and cubic spline interpolation respectively to repair the error frames.Comparing this method with BlazePose and lightweight-OpenPose,the results show that the overall recognition accuracy of this method is 3.1%and 7.8%higher than BlazePose and lightweight-OpenPose,respectively,and the detection of the wrist and elbow joints of the swing arm has been greatly improved.The execution efficiency is equivalent to BlazePose,which can assist the daily training of tennis service.

关 键 词：网球发球 人体关节识别 错误帧鉴别 综合修复 

分 类 号：G845[文化科学—体育训练] TP391.41[文化科学—体育学]