CNN-LSTM在多模态人体动作识别中的应用研究

打开文本图片集

中图分类号：TN911.73-34；TP391 文献标识码：A 文章编号：1004-373X（2025）15-0078-05

Research on application of CNN-LSTM in multimodal human body actionrecognition

WANG Tong （FujianNormal University，Fuzhou 350117，China）

Abstract：Since thereare multipletypes of objects intheenvironment，and lightingconditionsandoclusion mayffectthe effectofactionrecogntion，amultimodalhumanbodyactionrecognitionmethodbasedonCNN-LSTMisproposed.Twokindsof multimodalbehaviordata，depthimagedataandresultantacelerationdatawhichcanreflethumanbodyactions，arecolleted andinpuintoCNNmodel.Afterconvolution，dowsamplingandotherprocesing，theskeletal jointfeaturemapandresultantaccelerationvariancefeature mapofthehumanbodyactiondepth imageareextractedasthemultimodalfeature samplesof human bodyactions，whicharetheninputintotheLSTmodel.Andthen，incombinationwiththetemporalcharacteristicsof human bodyactions，theafinetransformationrelationshipbetweenthecurrentmultimodalfeaturesamplesandactiontypesislearned. Inalltimesteps，theactionclassficationresultsoutputbytheLSTMmodeloutputgatearecolleced，avotingdecisionmethod isintroduced，andtheclasificationresultswithtemostocurrencsaetakenastheulimodalumanbodyactionegition results.Theexperimentalresultsshowthattheproposedmethodcanrecognizemultipletypesof humanbodyactionsaccurately in both normal and low light environments.

Keywords：CNN-LSTM;multimodal humanbodyaction;typerecognition;Kinectsensor;skeletal joint;inertialsensor;re sultantacceleration

0 引言

人体动作识别专注于解析和分类人体的运动、姿势及行为模式，旨在实现对人体动作的精准自动辨识与深刻理解[1-2]。（剩余5446字）