基于图卷积神经网络的室内穿墙无源目标检测算法

Indoor Through-The-Wall Passive Target Detection Algorithm Based on Graph Convolutional Neural Network

针对室内穿墙场景下目标状态差异而导致的信道状态信息(Channel State Information,CSI)功率谱密度在时序发生相应变化规律的不同,本文提出了一种基于图卷积神经网络(Graph Convolutional Neural,GCN)的室内穿墙无源目标检测算法.不同于传统的基于CSI的统计特征实现目标检测的相关系统,该算法从CSI的图域出发,基于CSI时频图特征构建得到GCN图结构后,使用可实现对复杂图中各节点进行分类的GCN作为分类器,提高了室内复杂环境下目标检测的性能.该方法在对原始CSI进行异常值去除和小波阈值去噪的基础上,利用短时傅里叶变换得到每个子载波上CSI幅值的时频图;然后根据各子载波CSI时频图特点,将存在能量的频率平均分为5个频段,再计算每个频段的平均功率谱密度,并在每个时序对其进行排序;最后基于对平均功率谱密度排序后各频段索引的变化规律构造GCN图,并将其邻接矩阵和特征矩阵输入GCN网络中进行训练,最终实现图节点特征与目标状态的一一映射.实验结果表明,在玻璃墙和砖墙场景下,本文提出的算法能够很好地刻画目标状态不同而导致的CSI功率谱密度变化规律的差异,且其平均检测准确率均高于现有的R-TTWD(Robust device-free Through-The-Wall Detection)和TWMD(The-Wall Moving Detection)目标检测算法.

According to variation laws of channel state information(CSI)power spectral density(PSD)in the timing series caused by different target states in indoor through-the-wall scenarios,this paper proposes a passive target detection al⁃gorithm based on graph convolutional neural(GCN).Different from the traditional correlation system for target detection based on CSI statistical features,this algorithm starts from the graph domain of CSI,constructs the GCN graph structure based on CSI time-frequency diagram,and uses the GCN that can classify the nodes in the complex graph as the classifier,which improves the performance of target detection in the indoor complex environment.Based on outlier removal and wavelet threshold denoising for original CSI information,it uses the short-time Fourier transform to obtain the time-frequen⁃cy diagram of the CSI amplitude on each subcarrier.Then,according to the characteristics of each subcarrier’s CSI timefrequency diagram,the total spectrum is divided into five frequency bands on average,and the average PSD of each frequen⁃cy band is calculated and sorted at every sample time.Finally,a GCN graph is constructed based on the variation law of the index of each frequency band after sorting the average PSD,and then its adjacency matrix and feature matrix are input into the GCN network for training,which can finally realize the one-to-one mapping between graph node features and target states.Experimental results show that under the scenarios of glass wall and brick wall,the proposed algorithm can essential⁃ly characterize the difference of CSI PSD change regularity caused by different target states;and its average detection accu⁃racy is higher than that of the existing R-TTWD(Robust device-free Through-The-Wall Detection)and TWMD(The-Wall Moving Detection)target detection algorithms.