基于神经网络以全身关节点坐标估算侧切变向动作下肢动力学参数

Estimation of Lower Limb Dynamics Parameters for Sidestepping Based on Artificial Neural Networks by the Coordinates of the Whole Body Joint Points

目的:应用神经网络模型以人体关节点坐标为输入变量,估算运动员侧切变向动作过程中地面反作用力与下肢关节力矩。方法:采用Motion红外高速动作捕捉系统和Kistler三维测力台同步采集71名男子足球运动员完成侧切动作过程中的运动学和动力学数据。建立神经网络模型,以人体18个关节点坐标作为输入变量估算髋、膝、踝关节力矩和地面反作用力。采用相关系数、均方根误差及其标准化值评价神经网络估算效度;采用统计参数映射分析神经网络估算与实际曲线之间的差异;采用配对样本t检验分析神经网络估算所得地面反作用力峰值与实际值之间的差异。结果:神经网络估算的动力学参数与实测值之间具有较好相关性(r=0.82~0.97)。神经网络估算所得动力学参数在各个方向上的误差具有显著差异(P<0.050),矢状面具有较大的相关系数以及较小的标准化均方根误差。估算的侧切动作支撑阶段动力学指标曲线仅在地面反作用力前后分量(10%~12%,P=0.011;93%~95%,P=0.015)和垂直分量(5%~15%,P<0.001)以及髋关节内外翻力矩(99%~100%,P=0.015)和内外旋力矩(1%,P=0.017)的很少部分与实际曲线之间具有显著差异。神经网络估算所得的3个方向地面反作用力峰值与实测值之间均无显著差异。结论:应用神经网络模型以全身关节点坐标估算的运动员侧切变向动作过程中下肢关节力矩和地面反作用力与实测值之间相关性高,误差较小,其中矢状面效果最好。建立的神经网络模型可用于非实验室环境下监控侧切变向动作中的损伤风险。

Objective:Applying artificial neural networks(ANNs)with human joint coordinates as input variables to estimate the ground reaction forces(GRF)and lower limb joint torque during the sidestepping of athletes.Methods:Motion infrared high-speed motion capture system and Kistler three-dimensional force platform(FP)were used to synchronously collect the kinematics and dynamics data of 71 male football players in the process of completing the sidestepping.18 landmarks coordinates on whole-body as inputs in ANNs were used to estimate GRF and lower joint moments.The correlation coefficient,root mean square error(RMSE)and normalized root mean square error(nRMSE)were used to evaluate the validity of ANNs estimation.Statistical parameter mapping was used to analyze the difference between the ANNs estimation and the actual curves.Results:There was a good correlation between the kinetics estimated by the ANNs and the measured values(r=0.82—0.97).The errors of the kinetics estimated by the ANNs were significantly different in all directions(P<0.050).The sagittal plane had a higher correlation coefficient and fewer nRMSE.Only the anterior-posterior(10%—12%,P=0.011;93%—95%,P=0.015)and vertical(5%—15%,P<0.001)GRF estimated by ANNs were significantly different with measurement curves,as well as the frontal(99%—100%,P=0.015)and transverse(1%,P=0.017)hip moments.There was no significant difference between the peak values of the GRF in the three directions estimated by the ANNs and the measured by FP.Conclusions:With body landmarks coordinates as inputs,the GRF and lower limb joint moments of soccer players during sidestepping could be effectively estimated by ANNs,especially in the sagittal plane,which could be used in a non-laboratory environment.