基于双流CNN-BiLSTM的毫米波雷达人体动作识别方法

Human Action Recognition Method for Millimeter Wave Radar Based on Dual-Stream CNN-BiLSTM

目前基于雷达的人体动作识别方法,大多是先对人体动作的回波信号进行多维快速傅里叶变换(FFT)得到距离、多普勒和角度等信息,构造各种数据谱图后再输入到神经网络中进行分类识别,数据预处理过程较为复杂。提出了一种双流卷积神经网络(CNN)与双向长短时记忆网络(BiLSTM)串联的毫米波雷达人体动作识别方法。首先对原始的雷达回波信号复数采样数据(I/Q)进行帧差处理,以消除静态干扰,并将其转换为幅度/相位(A/P)的数据格式;然后将帧差后的I/Q和A/P数据分别输入单流的CNN-BiLSTM网络,提取人体动作的空间和时间特征,最后进行双流网络的融合以增强特征的交互性,提高识别准确率。实验结果表明,该方法数据预处理简单,并充分利用了动作数据的帧间相关性,模型收敛快,识别准确率可以达到99%,是一种快速有效的人体动作识别方法。

Most of current radar-based human action recognition methods first perform multidimensional fast Fourier transform(FFT)on the echo signal of human action to obtain information such as distance,Doppler and angle,construct various data spectrograms and then input them into a neural network for classification and recognition,which is a complicated data pre-processing process.A millimeter wave radar human action recognition method is proposed based on a dual-stream convolutional neural network(CNN)bridged with a bi directional long short-term memory network(BiLSTM)(denoted as CNN-BiLSTM).The original complex radar echo signal(I/Q)is first frame-differenced to eliminate static interference and then converted to amplitude/phase(A/P)data format.Then the frame-differenced I/Q and A/P data are fed into the single-stream CNN-BiLSTM network respectively to extract spatial and temporal features of human actions,and finally the fusion of the dual-stream network is performed to enhance the feature interaction and improve the recognition ac curacy.Experimental results show that the proposed method is simple in data pre-processing and makes full use of the inter-frame corre lation of action data,and the model converges quickly and the recognition accuracy can reach 99%,demonstrating that the proposed method is a fast and effective human motion recognition method.