基于深度学习的无标记运动捕捉的关键技术及其在运动科学中的应用  

The key technologies and the application in sports science of marker-less motion capture based on deep learning

作　　者：冯茹 杨辰 孙晓乐 戎科[1] 刘卉[2] Feng Ru;Yang Chen;Sun Xiaole;Rong Ke;Liu Hui(School of Sports and Health,Nanjing Sport Institute,Nanjing,Jiangsu;China Institute of Sports and Health,Beijing Sport University,Beijing)

机构地区：[1]南京体育学院运动健康学院,江苏南京 [2]北京体育大学中国运动健康研究院,北京

出　　处：《运动科学与健康研究》2025年第2期53-60,共8页Research on Sports Science and Health

基　　金：国家体育总局科技创新项目(23KJCX032),2023年江苏省高校“青蓝工程”项目。

摘　　要：人体运动捕捉是运动科学中生物力学研究的重要方法,目前可以通过人工识别、基于标记和无标记三种方法实现对人体运动的捕捉。人工识别技术耗时长且具有主观性,基于标记的运动捕捉技术影响人体运动且测试环境受限,而无标记运动捕捉方法可有效解决以上问题。无标记运动捕捉方法起源于计算机视觉和机器学习,该方法可从图像和视频中提取人体姿态,关键技术主要包括轮廓提取、姿态识别和跟踪三个方面。目前已有大量公开数据库用于新的无标记运动捕捉方法的开发。现有的已发布的无标记运动捕捉系统较少,包括STAGE、Theia3D和FastMove等。已有部分研究将该方法用于运动科学和体育运动的动作技术分析中,但其获取的人体运动学参数可能尚不能满足运动科学研究的精度需求。相对于传统的运动捕捉方法,无标记运动捕捉方法更为便携且成本也更低,未来有望广泛应用于运动科学研究中。Motion capture is an important method of biomechanics research in sports science.Motion capture could be realized through three methods manual recognition,marker-based and marker-less.The methods of manual recognition is time consuming and subjective.The method of marker-based affects human motion and the test environment is limited,while marker-less motion capture method can effectively solve these problems.Marker-less motion capture originated from computer vision and machine learning,which could estimate human posture from images and videos.The key technologies mainly include contour extraction,pose estimation and tracking.Marker-less motion capture is under development.There are a large number of public datasets for the development of marker-less motion capture.A few marker-less motion capture systems were published,including STAGE,Theia3D,FastMove and so on.Some studies have applied marker-less motion capture to the research of sports science and motion analysis,but the kinematic parameters might not meet the accuracy requirements of sports science.Compared with traditional motion capture methods,marker-less motion capture method is more portable and lower cost.Marker-less motion capture would be widely used in sports science research in the future.

关 键 词：深度学习 无标记运动捕捉 数据集 运动科学 动作分析 

分 类 号：G804.6[文化科学—运动人体科学]