【作者】 王平；

【导师】 陶智勇； 凌毓；

【作者基本信息】 武汉邮电科学研究院 ， 通信与信息系统， 2022， 硕士

【摘要】 由于雷达设备具有不接触性和信号保密性高等特点,使得雷达设备被广泛应用在居家养老、探测、搜救等领域,因此应用环境对雷达信号的处理提出了很高的要求,传统的信号时频滤波和特征提取,需要从时域变换到频域,再反变换到时域,这样的流程不利于并行操作,存在时间的滞后,本文提出了一种基于超宽带雷达的双通道输入跌倒信号检测方法,避开复杂的信号处理过程,只在回波信号时域的通道上做一次快速傅里叶变换和一次奇异值分解,在频域滤波时域重构之后,就可以得到超宽带雷达回波的时域特征图和频域特征图,通过搭建的一个双通道输入网络模型就可以实现对跌倒动作的检测,并且数据在时域重构的同时,做到了数据压缩,减小了数据存储空间,而且时频操作互不影响,可以同时运行。本文主要工作如下:(1)基于一种常规的超宽带(UWB)雷达,将其部署到模拟环境中对跌倒动作进行了数据采集,收集并建立了一个跌倒和非跌倒动作样本数据集,共444条样本,该数据集共包含9名不同身高、不同体重的参与者的3种不同的跌倒动作样本共422条,非跌倒数据22条。(2)基于(1)中的数据,根据雷达数据特性本论文提出了一种时频分别处理的方法,具体为在快时间通道做一次快速傅里叶变换,在全局做一次奇异值分解重构,筛选出雷达回波的时频特征,同时提出了基于能量阈值的数据分割方法,切分出最有价值的时频数据。(3)在时频特征提取的基础上,本论文设计了一个时频双输入通道的网络跌倒监测模型来判别三种跌倒行为方式,通过与其他常规机器学习方法以及单一通道输入比较,验证了该方法的稳定性和准确性,所提的方法对三类跌倒检测的准确率到达94.92%。更多还原

【Abstract】 In view of the characteristics of non-contact and signal confidentiality of radar equipment,radar equipment is widely used in the fields of home care,detection,search and rescue,etc.Therefore,high requirements are placed on the processing of radar signals.The traditional signal time-frequency filtering and feature extraction need to be transformed from time domain to frequency domain,and then inversely transformed to time domain.Such a process is not conducive to parallelization and has time lag.This thesis proposes an ultra-wideband radar fall signal based on convolutional neural network monitoring method.It avoids the complicated signal processing process,and only performs a Short-time Fourier Transform and a Singular Value Decomposition on the time-domain channel of the echo signal to obtain the frequency domain map and the time domain map,and return the ultra-wideband radar the time-domain image and frequency-domain image of the wave are used as two channels,and the method of convolutional neural network in target recognition is used to monitor the fall action,and good results are obtained.The main work of this thesis is as follows:(1)A conventional Ultra-wide Band(UWB)radar was deployed in a simulated environment to collect data on falling movements.A sample dataset of falling and non-falling movements was collected and established,which included 3 different movement types of 9participants of different heights and weights.(2)Based on the above data,a time-frequency joint filtering algorithm is proposed according to the radar echo signal data of human body.Through the measurement experiment,the time-frequency single filtering method and the joint filtering method are compared to select the joint filtering method and screen out the main motion features.(3)Based on the above action feature graph,this thesis designs a fall monitoring model based on convolutional neural network to judge the fall behavior.This model has better stability and accuracy for the discrimination effect of UWB radar data containing action features.Experimental results show that the accuracy of the proposed method can reach 94.92%.更多还原