第36卷第2期2023年2月传感技术学报CHINESEJOUＲNALOFSENSOＲSANDACTUATOＲSVol.36No.2Feb．2023项目来源:浙江省基础公益研究计划项目(LGF21F020017)收稿日期:2022－04－12修改日期:2022－05－14ALightweightConvolutionalNeuralNetworkFallDetectionAlgorithmUsingInertialSensors*LIUPengda1，PANJulong1*，ZUOZhengwei2，ZHUHailiang1，LIYanli1(1．CollegeofInformationEngineering，ChinaJiliangUniversity，HangzhouZhejiang310018，China;2．ModernEducationTechnologyCenter，ChinaJiliangUniversity，HangzhouZhejiang310018，China)Abstract:Deployingatinymachinelearning(TinyML)modelinawearablefalldetectionterminalhasproblemssuchasweakcomputingpower，limitedmemory，andincompletemanualfeatureselectionthroughtraditionalmachinelearningalgorithms．Alightweightconvolu-tionalneuralnetworkfalldetectionalgorithmusinginertialsensersisproposed，andahigh-precisionwearablefalldetectionsystemisdesignedandimplemented．Thealgorithmautomaticallyextractsmorecompletedatafeaturesfromthefalldataandusesthedepthwiseseparableconvolutiontodecomposethestandardconvolutionintodepthwiseconvolutionandpointwiseconvolution．Withonly0．2%ofthefalldetectionaccuracylost，theamountoflayerparametersisreducedby75．32%，makingitmoresuitableforbeingdeployedinre-source-constrainedembeddedterminals．Theexperimentalresultsshowthatthealgorithmachievestheaverageaccuracy，sensitivityandspecificityof99．29%，98．00%and100．00%，respectively，intheactualfalltestenvironment．Comparedwithotheralgorithms，thisalgo-rithmnotonlyreducesthemodelsizeandcalculationamount，butalsoensuresthedetectionaccuracyoffalldetection．Thesuccessfuldevelopmentofthissystemprovidesanewwayforthefalldetectionandalarmoftheelderly．Keywords:falldetection;convolutionalneuralnetwork;TinyML;inertialsensors;deeplearningEEACC:7230doi:10．3969/j．issn．1004－1699．2023．02．013基于惯性传感器的轻量化卷积神经网络跌倒检测算法*刘鹏达1，潘巨龙1*，左正魏2，朱海亮1，李艳丽1(1．中国计量大学信息工程学院，浙江杭州310018;2．中国计量大学现代教育技术中心，浙江杭州310018)摘要:在可穿戴式跌倒检...
