基于车载毫米波雷达动态手势识别网络董连飞，马志雄，朱西产（同济大学汽车学院智能汽车研究所,上海201804）摘要：基于Transformer提出一种车载毫米波雷达手势识别方法，可用于车内人员进行人机交互.毫米波雷达信号不受车内光照变化影响，同时能够保证乘客隐私.首先，毫米波雷达采样信号经过二维傅里叶变换和滤波获得距离—多普勒（RDM）和距离—角度图(RAM)；然后，将连续多帧RDM和RAM经过三维卷积网络后进行特征融合与拼接得到特征向量，利用Transformer模块进行位置和序列编码；最后通过全连接层获得手势概率分布向量.采集了基于实际路况和多种干扰环境下的雷达数据制作了手势识别分类的数据集，实验结果表明该方法可以有效的检测与识别多种典型手势，识别准确率可以达到97.14%以上.关键词：动态手势识别；三维卷积神经网络；毫米波雷达中图分类号：TN95文献标志码：A文章编号：1001-0645(2023)05-0493-06DOI：10.15918/j.tbit1001-0645.2022.102DynamicGestureRecognitionNetworkBasedonVehicularMillimeterWaveRadarDONGLianfei，MAZhixiong，ZHUXichan（IntelligentVehicleResearchInstitute,SchoolofAutomotiveStudies,TongjiUniversity,Shanghai201804,China）Abstract：ATransformerbasedmillimeterwaveradargesturerecognitionmethodwasproposedforhuman-com-puterinteractionofvehicleoccupants.Themillimeterwaveradarsignalwasdesignedtobenotaffectedbythechangeoflightinsidethevehicle,andatthesametimetoensuretheprivacyofpassengers.Firstly,themilli-meterwaveradarsampledsignalwascarriedthroughtwo-dimensionalFouriertransformandfilteringtoobtaindistance-Doppler(RDM)anddistance-anglemaps(RAM).Then,consecutivemulti-frameRDMandRAMwerefusedandstitchedafterthree-dimensionalconvolutionalnetworkstoobtainfeaturevectors.AndaTransformermodulewasusedtoperformpositionandsequenceencoding.Finally,thegestureprobabilitydistributionvectorwasobtainedthroughthefullyconnectedlayer.Adatasetforgesturerecognitionclassificationwascollectedbasedontheactualroadconditionsandradardataunderavarietyofinterferenceenvironments.Theexperiment-alresultsshowthatthemethodcaneffectivelydetectandrecognizeavarietyoftypicalhandgestures,andthere-cognitionaccuracycanreachmorethan97.14%.Keywords：dynami...
