无反光点人体运动自动捕捉人工智能系统的有效性  

Validity of an Artificial Intelligence System for Markerless Human Movement Automatic Capture

作　　者：刘卉[1] 李翰君[1] 曲毅 何晓光 周志鹏[3] 于冰 LIU Hui;LI Han-jun;QU Yi;HE Xiao-guang;ZHOU Zhi-peng;YU Bing(Beijing Sport University,Beijing 100084,China;Dalian Ruidong Technology Co.,Ltd.,Dalian 116033,Liaoning China;Shandong Sport University,Jinan 250102,Shandong China;Chinese Athletics Association,Beijing 100763,China)

机构地区：[1]北京体育大学,北京100084 [2]大连锐动科技有限公司,辽宁大连116033 [3]山东体育学院,山东济南250102 [4]中国田径协会,北京100763

出　　处：《北京体育大学学报》2021年第1期125-133,共9页Journal of Beijing Sport University

基　　金：科技部重点研发计划项目“冬季体能类项目主要训练手段功效的研究”(项目编号:2018YFF0300404);中国田径协会备战东京奥运会科技助力服务项目“国家田径队重点投掷运动员实战比赛技术的综合分析与评价”“田径投掷项目国家队备战东京奥运会(2020年)重点运动员比赛技术分析与评定科技服务”(项目编号:2019TJ018,2020TJ02008)。

摘　　要：北京体育大学联合大连锐动科技有限公司和中国田径协会研发了一套基于机器学习的人体运动自动捕捉人工智能系统。目的:检验使用人体运动自动捕捉人工智能系统获得的人体关节点三维坐标的有效性。方法:在实际比赛中使用2台录像机拍摄12名女子标枪运动员的试投,取各自最好试投成绩的视频进行分析。使用直接线性转换(DLT)方法对录像机进行标定,标定误差=(0.003 4±0.001 2)m。4名有多年经验的研究人员独立手工解析动作视频中人体21个关节点的二维坐标,并合成三维坐标。同时使用人体运动自动捕捉人工智能系统自动识别了每次试投中21个关节点的二维坐标,并合成三维坐标(自动解析曲线)。计算每名运动员每个关节点4名研究人员手工解析合成的三维坐标的平均值,作为关节点三维坐标最佳值(手工解析平均曲线)。计算自动解析曲线和手工解析平均曲线之间的多重相关系数和差值的平均值、标准差和95%置信区间。结果:除了右脚尖Z坐标因为坐标变化量比较小外,自动解析曲线与手工解析平均曲线之间的多重相关系数均大于0.95,85%的自动解析曲线与相应的手工解析平均曲线的多重相关系数大于0.98。自动解析曲线与手工解析平均曲线之间的差值全部小于0.015 m,50%的自动解析曲线与手工解析平均曲线之间的差值小于0.010 m。结论:基于机器学习的人体运动自动捕捉人工智能系统准确地模拟了手工解析过程,所获得的人体关节点三维坐标与有经验的研究人员手工解析获得的三维坐标平均值高度相似。该系统极大地节省了科技助力的人力并提高了数据反馈速度,已成功应用在多个田径项目和冬季运动项目的科技助力工作中。Beijing Sport University in collaboration with FastMove Technology Inc,and Chinese Athletics Association recently developed an artificial intelligence system for markerless human movement automatic capture.The purpose of this study was to determine the validity of the three-dimensional(3 D)coordinate data collected by this markereless human movement automatic capture system.Twelve female javelin throwers’ performance in actual competition were recorded using two video camcorders.Camcorders were calibrated using Direct Linear Transformation(DLT)method.The resultant calibration error was 0.003 4±0.001 2 m.The best trial of each thrower was used in this study.The video records of each trail were manually digitized by four individuals with rich experience in manual digitizing,and also by human movement automatic capture system.Three-dimensional coordinates of 21 body landmarks were obtained.Four 3 D coordinate data sets of each trial obtained through manual digitizing were averaged to obtain the mean manually digitized data of the trial.The coefficients of multiple correlation(CMCs)and absolute errors(AEs)of 21 body landmarks between automatically captured 3 D coordinate data and the corresponding mean data obtained through manual digitizing were reduced.The means,standard deviations,and 95%confidence intervals of CMCs and AEs were calculated.Except the Z coordinate right toe,CMCs of the coordinates of all body landmarks were greater than 0.95 with 85% of them greater than 0.98.The CMC of the Z coordinate of the right toe was below 0.95 because the range of movement was small.AEs of the coordinates of all body landmarks were less than 0.015 m with 50% of them less than 0.010 m.The new human movement automatic capture system were applied in several disciplines of track and field and winter sports.The new human movement automatic capture system significantly increased the speed of biomechanical feedbacks.The results of this study suggest that new human movement automatic capture system based on artificial intelligen

关 键 词：人工智能 动作捕捉 录像分析 关节坐标 生物力学 

分 类 号：G804.6[文化科学—运动人体科学]