基于体感识别技术的运动训练辅助系统设计  

Design of sports training assistant system based on somatosensory recognition technology

作　　者：郑永权 张飞云[1] 周帅[1] ZHENG Yongquan;ZHANG Feiyun;ZHOU Shuai(Xi’an Jiaotong University City College,Xi’an 710018,China)

机构地区：[1]西安交通大学城市学院,陕西西安710018

出　　处：《电子设计工程》2021年第24期93-97,共5页Electronic Design Engineering

基　　金：陕西省体育局常规课题(2020031)。

摘　　要：针对传统运动训练方式严重依赖教练人工指导的问题,文中将体感识别技术应用于运动训练辅助系统的设计中,以实现运动训练智能化。该运动训练辅助系统采用Kinect V2作为实时动作采集传感器,使用分隔策略将人物从运动场所背景中分离出来,从而降低数据的计算量。通过将人体简化为18个骨骼关节降低动作识别的复杂度,并使用多目标跟踪算法来捕捉关节点位置数据。利用VGG卷积神经网络将二维关节数据转换成人体姿态图,将不同时刻的人体姿态图作为堆叠模型的训练样本,以监督学习的方式进行训练和参数优化,得到运动训练实时动作识别模型。经过测试和数据分析证明,文中所提的系统设计方案具有较好的鲁棒性与实用性。Aiming at the problem that traditional sports training methods rely heavily on the manual guidance of coaches,the article applies somatosensory recognition technology to the design of sports training auxiliary systems to realize sports training intelligence.The sports training assistance system uses Kinect V2 as a real⁃time action acquisition sensor,and uses a separation strategy to separate the characters from the background of the sports venue,thereby reducing the amount of data calculation.By simplifying the human body to 18 skeletal joints,the complexity of motion recognition is further reduced,and the multi⁃target tracking algorithm is used to capture the position data of joint points.Use the VGG convolutional neural network to generate a human body pose map from the two⁃dimensional joint data,use the human body pose maps at different moments as the training samples of the stacked model,and perform training and parameter optimization in a supervised learning manner,and obtain a real⁃time motion recognition model for sports training.Tests and data analysis prove that the system design proposed in the article has good robustness and practicality.

关 键 词：KINECT 分隔策略 多目标跟踪算法 VGG卷积神经网络 堆叠模型 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]