周成龙,
介钧誉,
周涛,
汪相锋,
徐丰,
复旦大学电磁波信息科学教育部重点实验室 上海 200433
基金项目:国家自然科学基金(61822107)
详细信息
作者简介:夏朝阳:男,1993年生,博士生,研究方向为雷达信号处理、目标识别和深度学习
周成龙:男,1995年生,硕士生,研究方向为雷达目标识别和深度学习应用
介钧誉:男,1993年生,硕士生,研究方向为雷达目标识别和深度学习应用
周涛:男,1996年生,硕士生,研究方向为雷达目标识别和深度学习应用
汪相锋:男,1995年生,硕士生,研究方向为雷达目标识别和深度学习应用
徐丰:男,1982年生,教授,博士生导师,研究方向为SAR图像解译、电磁散射建模和类脑人工智能
通讯作者:徐丰 fengxu@fudan.edu.cn
中图分类号:TN958.94计量
文章访问数:2664
HTML全文浏览量:890
PDF下载量:176
被引次数:0
出版历程
收稿日期:2019-10-16
修回日期:2019-11-27
网络出版日期:2019-12-09
刊出日期:2020-01-21
Micro-motion Gesture Recognition Based on Multi-channel Frequency Modulated Continuous Wave Millimeter Wave Radar
Zhaoyang XIA,Chenglong ZHOU,
Junyu JIE,
Tao ZHOU,
Xiangfeng WANG,
Feng XU,
Key Laboratory for Information Science of Electromagnetic Waves(MoE), Fudan University, Shanghai 200433, China
Funds:The National Natural Science Foundation of China (61822107)
摘要
摘要:该文提出一种基于多通道调频连续波(FMCW)毫米波雷达的微动手势识别方法,并给出一种微动手势特征提取的最优雷达参数设计准则。通过对手部反射的雷达回波进行时频分析处理,估计目标的距离多普勒谱、距离谱、多普勒谱和水平方向角度谱。设计固定帧时间长度拼接的距离-多普勒-时间图特征,与距离-时间特征、多普勒-时间特征、水平方向角度-时间图特征和三者联合特征等,分别对7类微动手势进行表征。根据手势运动过程振幅和速度差异,进行手势特征捕获和对齐。利用仅有5层的轻量化卷积神经网络对微动手势特征进行分类。实验结果表明,相较其他特征,设计的距离-多普勒-时间图特征能够更为准确地表征微动手势,且对未经训练的测试对象具有更好的泛化能力。
关键词:毫米波雷达/
微动手势识别/
调频连续波/
卷积神经网络
Abstract:A micro-motion gesture recognition method based on multi-channel Frequency Modulated Continuous Wave (FMCW) millimeter wave radar is proposed, and an optimal radar parameter design criterion for feature extraction of micro-motion gestures is presented. The time-frequency analysis process is performed on the radar echo reflected by the hand, and the range Doppler spectrum, the range spectrum, the Doppler spectrum and the horizontal direction angle spectrum of the target are estimated. Then the range-Doppler-time-map feature is designed, range-time-map feature, Doppler-time-map feature, horizontal-angle-time-map feature, and three-joint feature with fixed frame time length are used to characterize the 7 classes micro-motion gestures, respectively. And these gesture features are captured and aligned according to the difference in amplitude and speed of the gesture motion process. Then a five-layer lightweight convolutional neural network is designed to classify the gesture features. The experimental results show that, the range-Doppler-time-map feature designed in this paper characterizes the micro-motion gesture more accurately and has a better generalization ability for untrained test objects compared with other features.
Key words:Millimeter wave radar/
Micro-motion gesture recognition/
Frequency Modulated Continuous Wave(FMCW)/
Convolutional Neural Network(CNN)
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=5667621c-2c6a-484d-b3c3-13c2a64d16ae