基于RDATM数据集的毫米波雷达动态手势识别系统设计

Dynamic Gesture Recognition with FMCW Radar Based on RDATM Dataset

针对现有毫米波雷达手势识别研究中手势信息提取不充分的问题,提出一种基于RDATM三维参数数据集的手势识别方法。该方法先对获取的雷达手势回波数据进行预处理并进行动态手势检测,然后利用回波信号频谱信息构造距离-时间图(Range-Time Map, RTM),多普勒-时间图(Doppler-Time Map, DTM)和角度-时间图(Angle-Time Map, ATM),最后利用RTM、DTM、ATM构建RDATM数据集并送入基于VGG16网络改进的单分支卷积神经网络进行手势特征提取和分类。实验结果表明,利用RDATM数据集对6种手势的平均分类准确率可高达99.17%。且文中改进的网络相较于原VGG16模型对同类型数据集训练时间更短、分类准确率更高。

Aiming at the problem of insufficient gesture information extraction in existing FMCW radar gesture recognition research, a gesture recognition method based on RDATM three-dimensional parameter dataset is proposed. Firstly, the obtained radar gesture echo data is preprocessed and conduct dynamic gesture detection. Then, construct Range-Time-Map(RTM), Doppler-Time-Map(DTM) and Angle-Time-Map(ATM) by analysing spectrum information of echo signal. Finally, RDATM dataset is established by fusing RTM, DTM, ATM and sent it to the single branch convolutional neural network based on improved vgg16 network for gesture feature extraction and classification. The experimental results show that the average classification accuracy of six gestures can be as high as 99.17%. Compared with the original vgg16 model, the improved network has shorter training time and higher classification accuracy for the same type of datasets.