基于时空图注意力网络的跳高轨迹预测  

High Jump Trajectory Prediction Based on Spatial-temporal Graph Attention Network

作　　者：钟亦强 尚维[1] 石坚[1] 程燕[3] 苑廷刚[3] ZHONG Yi-qiang;SHANG Wei;SHI Jian;CHENG Yan;YUAN Ting-gang(Academy of Mathematics and System Science,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100190,China;Institute of Sports Science,General Administration of Sport of China,Beijing 100061,China)

机构地区：[1]中国科学院数学与系统科学研究院,北京100190 [2]中国科学院大学管理决策与信息系统实验室,北京100190 [3]国家体育总局体育科学研究所,北京100061

出　　处：《数学的实践与认识》2022年第5期64-73,共10页Mathematics in Practice and Theory

基　　金：国家体育总局体育科学研究所(基本20-35课题)。

摘　　要：跳高运动由助跑、单脚起跳、越过横杆落地等一系列动作组成,跳高运动轨迹是一个复杂的因果系统.在每一个时间节点,运动员的身体姿态符合一定的规律约束.研究建立了跳高运动过程中关键身体关节节点坐标依时间变化的离散过程模型,每个时刻的各关节点位置向量由给定的之前若干个时刻的运动员关节点位置决定.跳高运动员肢体的摆动、关节点位置的变化都对运动员位移有重要作用.运动员重要关节点之间的相互作用,使得跳高运动轨迹难以刻画和预测.随着深度学习的发展,图神经网络在时空特征刻画和预测上具有优势.基于时空图注意力网络结合长短时记忆网络、门控循环单元(STGAT-LSTM、STGAT-GRU)建立了具有空间位置约束的跳高运动员关节点位置随时间变化的模型,用于推算跳高运动员重心和关节点位置轨迹.实验结果表明该模型能够很好地刻画跳高运动的过程,对于关节点位置的预测能够达到较好的效果.研究所提出模型的预测效果比基线模型准确性提高了0.02-0.03米.研究给出了基于神经网络模型进行跳高运动轨迹预测的可行解决方案,可作为跳高运动仿真和技术优化研究的重要参考.The high jump movement consists of a series of movements such as running up,taking off on one foot,landing over the bar and so on.The track of the high jump movement is a complex cause-and-effect system.At each time point,the body posture of athletes conform to certain rules and constraints.In this study,the discrete process model of key body joint node coordinates in the process of high jump is established,and the position vector of each joint node at each moment is determined by the given position of the athlete ’s joint node at several previous moments.The swing of the high jumper’s body and the change of the position of the joint play an important role in the displacement of the high jumper.The interaction between the important points of the athletes makes the trajectory of the high jump difficult to be characterized and predicted.With the development of Deep learning,Graph Neural Network has advantages in characterization and prediction of spatio-temporal capture.Based on Graph attention network combined with Long and Short Time Memory network(LSTM) and Gated Recurrent Units(STGAT-LSTM and STGAT-GRU),we established a spatio-temporal model of high jump athletes’ joint position changing with time,which was used to calculate the center of gravity and joint position trajectory of high jump athletes.The experimental results show that the model can describe the process of high jump and predict the position of joints well.In this study,the prediction effect of the proposed model is 0.02-0.03 meters higher than that of the baseline model.In this study,a feasible solution for high jump trajectory prediction based on neural network model is given,which can be used as an important reference for high jump simulation and technical optimization research.

关 键 词：图注意力网络 长短时记忆网络 跳高轨迹位置预测 时空信息 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] G823.1[自动化与计算机技术—控制科学与工程]