基于数据增强的实时人体动作识别

打开文本图片集

中图分类号：TP391.4 文献标志码：A 文章编号：2095-2945（2025）18-0009-08

Abstract：Humanactionrecognition（HAR）basedonChannel State Information（CSI）hassignificantappicationprospectsin feldssuchashuman-computerinteraction，healthcare，andintrusiondetection.Althoughcurentresearchhasmadesubstantial progressinrecognizingvarioustypesofactivitesandimprovingrecognitionaccracy，challengesremainintheneedforalarge numberofactivitysamplestotrainmodels，andinimprovingthereal-timeperformaneoftherecognitionprocessToaddress theseissues，areal-time humanactionrecognition（HAR）system，CSI-FHAR，isdesignedbasedondataaugmentation.By augmentingasmallnumberofreal samples to generatesyntheticsamples，thesystemreduces thedemandforreal samples during modeltraning.Aditioallbysegmentingcompleteactivitysamples，therecogitionspeedisincreased，ehancingal-ie performance.Toincreasetheinter-classfeaturediferences，CSI-FHARaddstemporalencodingtotheactivitysamples，tereby improvingthemodelsrecognitionacuracy.Theconvolutionalblockatentionmodule（CBAM）isembeddedintheconvolutional neuralnetwork（CNN）tofurtherenhancethenetwork'sabilitytoextractefectivefeaturesfromactivitysamples.Experimental resultsemonstratetheefectivenessofCSI-FHAR：withonlyfivesamplesperactivityclassfor1typesofactites，the proposed model achieved recognition accuracies of 95.1% for gestures and 92.5% for full-body activities.

Keywords：humanactionrecognition（HAR）；channelstateinformation;dataenhancement;temporalcoding；aention mechanism

人体动作识别（Human Activity Recognition，HAR）技术在人机交互、医疗健康、智能家居和日常行为检测等领域获得了广阔的应用。（剩余10755字）