2023⁃03⁃10计算机应用,JournalofComputerApplications2023,43(3):972-977ISSN1001⁃9081CODENJYIIDUhttp://www.joca.cn基于人体模型约束的步态动态识别方法刘今越*,李慧宇,贾晓辉,李佳蕊(河北工业大学机械工程学院,天津300401)(∗通信作者电子邮箱ljy@hebut.edu.cn)摘要:针对外骨骼机器人在人机交互、医疗康复中的人体运动步态准确识别问题,提出一种基于人体模型约束的步态动态识别方法。首先,利用AMS仿真软件建立不同运动的仿真模型,根据模型约束划分步态相位,并通过回归映射建立真实数据与仿真数据间的对应关系;然后,将柔性压力传感器采集的足底压力数据以及惯性测量单元采集的足部位移数据融合为足部运动数据,并根据动态变化结合模型约束条件动态分割运动数据,以判断步态相位;最后,搭建卷积神经网络(CNN)识别行走步态相位。实验结果表明,所提方法的行走动作步态平均识别准确率为94.58%,上、下楼梯动作的平均步态识别准确率分别为93.21%和94.64%,与未经动态分割的足底压力数据的步态识别相比,分别提高了11.34、12.19和16.03个百分点。可见,通过经动态分割的足部运动数据进行CNN识别具有较高的准确率,且适用于不同动作的步态识别。关键词:步态识别;动态检测;人体模型;卷积神经网络;足底压力中图分类号:TP242.6文献标志码:ADynamicgaitrecognitionmethodbasedonhumanmodelconstraintsLIUJinyue*,LIHuiyu,JIAXiaohui,LIJiarui(SchoolofMechanicalEngineering,HebeiUniversityofTechnology,Tianjin300401,China)Abstract:Aimingattheissueofaccuraterecognitionofhumanmotiongaitinexoskeletonrobothumancomputerinteractionandmedicalrehabilitation,adynamicgaitrecognitionmethodbasedonhumanmodelconstraintswasproposed.Firstly,AnybodyModelingSystem(AMS)simulationsoftwarewasusedtoestablishdifferentmotionsimulationmodels,thegaitphasesweredevidedaccordingtothemodelconstraints,andthecorrespondingrelationshipbetweentherealdataandthesimulationdatawasestablishedthroughregressionmapping.Then,theplantarpressuredatacollectedbytheflexiblepressuresensorandthefootdisplacementdatacollectedbytheinertialmeasurementunitwerefusedintothefootmotiondata,andthemotiondatawasdynamicallysegmentedaccordingtoitsdynamicchangesandthemodelconstraintstodeterminethegai...