北京航空航天大学学报 ›› 2018, Vol. 44 ›› Issue (6): 1117-1123.doi: 10.13700/j.bh.1001-5965.2017.0397

• 论文 • 上一篇    下一篇

基于卷积神经网络的手势动作雷达识别方法

王俊, 郑彤, 雷鹏, 张原, 樵明朗   

  1. 北京航空航天大学 电子信息工程学院, 北京 100083
  • 收稿日期:2017-06-12 出版日期:2018-06-20 发布日期:2018-06-28
  • 通讯作者: 雷鹏.E-mail:peng.lei@buaa.edu.cn E-mail:peng.lei@buaa.edu.cn
  • 作者简介:王俊 男,博士,教授,博士生导师。主要研究方向:雷达信号处理、实时信号处理;郑彤 女,博士研究生。主要研究方向:信号处理、模式识别;雷鹏 男,博士,讲师。主要研究方向:信号处理、模式识别。
  • 基金资助:
    国家自然科学基金(61501011,61671035)

Hand gesture recognition method by radar based on convolutional neural network

WANG Jun, ZHENG Tong, LEI Peng, ZHANG Yuan, QIAO Minglang   

  1. School of Electronic and Information Engineering, Beijing University of Aeronautics and Astronautics, Beijing 100083, China
  • Received:2017-06-12 Online:2018-06-20 Published:2018-06-28

摘要: 随着手势动作识别技术在人机交互、生活娱乐及医疗服务等应用领域的逐步深入,其对非接触、微光条件下的稳健测量与识别能力提出更高要求。针对该问题,研究了一种基于线性调频连续波(LFMCW)雷达距离-多普勒(RD)信息和卷积神经网络(CNN)的典型手势动作识别方法。首先,对于LFMCW雷达回波,通过去斜、快时间域快速傅里叶变换和相干积累,获取手势目标的二维RD像数据;其次,以RD像幅度矩阵作为CNN输入样本,利用2层卷积与池化处理构建特征空间,从而通过全连接与softmax分类器实现对手势动作的有效识别;最后,在此基础上,采用24 GHz工业雷达传感器设计手势测量实验系统,形成关于4种典型手势动作的LFMCW雷达回波数据库。实验结果表明,将24 GHz LFMCW雷达回波RD处理与CNN结合能够实现对典型手势动作的有效识别。

关键词: 手势动作识别, 线性调频连续波(LFMCW)雷达, 距离-多普勒(RD), 卷积神经网络(CNN), softmax分类器

Abstract: With the widespread use of hand gesture recognition technique, capabilities of robust measurement and classification in non-contact and all-day conditions are much desired in its applications, such as human-computer interaction, life entertainment and medical service. According to this requirement, the paper introduces a hand gesture recognition method based on linear frequency modulated continuous wave (LFMCW) radar range-Doppler (RD) information and convolutional neural network (CNN). Firstly, for LFMCW radar echoes from hand gestures, dechirping, fast Fourier transform in fast-time domain and coherent integration are applied to produce the two-dimensional RD images of hand gesture. Next, they are used as the input data of CNN, and the feature space is constructed with the process of two-layer convolution and pooling. Finally, the effective hand gesture recognition is achieved by full connection and softmax classifier. On this basis, a 24 GHz industrial radar sensor is used to design the experimental system for hand gesture measurement, and a dataset of four typical hand gestures is also generated with the LFMCW waveform. The experimental results show that the proposed method based on RD information and CNN is applicable to general radar sensors at 24 GHz and could achieve effective recognition of typical hand gestures.

Key words: hand gesture recognition, linear frequency modulated continuous wave (LFMCW) radar, range-Doppler (RD), convolutional neural network (CNN), softmax classifier

中图分类号: 


版权所有 © 《北京航空航天大学学报》编辑部
通讯地址:北京市海淀区学院路37号 北京航空航天大学学报编辑部 邮编:100191 E-mail:jbuaa@buaa.edu.cn
本系统由北京玛格泰克科技发展有限公司设计开发