基于视频的上肢外骨骼行为预判方法  

Video based behavior prediction method of upper limb exoskeleton

作　　者：马六章 蒋磊[1] 吴越 程子均 MA Liu-zhang;JIANG Lei;WU Yue;CHENG Zi-jun(School of Mechanical Electronic and Information Engineering,China University of Mining and Technology-Beijing,Beijing 100083,China)

机构地区：[1]中国矿业大学(北京)机电与信息工程学院,北京100083

出　　处：《计算机工程与设计》2022年第5期1419-1427,共9页Computer Engineering and Design

基　　金：国家自然科学基金重点基金项目(61936008)。

摘　　要：针对传统穿戴式上肢外骨骼人机交互行为预测受到穿戴者身体状况影响的问题,提出一种基于卷积神经网络(CNN)的多时间融合(multiple temporal fusion,MTF)模块。对特征进行一系列的子卷积处理,使每一帧可以完成多个邻域的时间聚合,能够在距离较远的帧上建立长期的时间关系,在推理时提出一种基于非线性最小二乘参数的自适应算法(NRLS-A),对神经网络参数进行实时调整,使在线预测精度提高了28%。该模型基于TensorFlow在GPU上进行并行计算。在自建数据集上的平均准确度达到了84.9%,比传统的LRCN网络模型提高了5%。Aiming at the problem that the traditional wearable upper limb exoskeleton human-computer interaction behavior prediction was affected by the wearer’s physical condition,a multi temporal fusion(MTF)module based on convolutional neural network(CNN)was proposed.A series of sub convolution processing was carried out on the features,so that each frame could complete the time aggregation of multiple neighborhoods,and a long-term time relationship on the far away frames was established.An adaptive algorithm based on nonlinear least squares parameters(NRLS-A)was proposed in the reasoning,which adjusted the neural network parameters in real time,so that the online prediction accuracy was improved by 28%.The model was based on tensorflow to perform parallel computing on GPU.The average accuracy on the self-built data set is 84.9%,which is 5%higher than the traditional LRCN network model.

关 键 词：上肢外骨骼 视频预测 卷积神经网络 主成分分析 门控循环网络 非线性最小二乘 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]