Radar Vital Signs Detection by Using Optimized CNN+GRU Model

Abstract

Wearable devices may have significant disadvantages compared to non-contact sensors. In addition, heart and respiratory rate information may be insufficient for expert evaluation. Radar sensors allow vital signals to be detected without interfering with subject activities. In this study, a high performance Convolutional Neural Network (CNN) model is proposed to classify four different types of radar vital signs. Firstly, raw radar data is converted into 2-dimensional matrix form by spectrogram method in order to analyze raw radar data in time and frequency environment. Spectrogram processing on raw radar data is required for training and testing of pre-trained CNN networks. Afterwards, high-level features obtained from pre-trained CNN networks were fused by using Canonical Correlation Analysis (CCA). These features were used to train and test for Gated Recurrent Unit (GRU) block structure optimized by Whale Optimization algorithm (WOA). The proposed Spectrogram Content Based Optimized CNN + GRU Model performance is 95.52% Accuracy (ACC), 94.74% Sensitivity (SEN), 98.29% Specificity (SPE=, 96.35% Precision (PRE), 95.40% F1-Score and 93.96% Matthews Correlation Coefficient (MCC). Our proposed method showed that preferability potential of non-destructive radar technology is more valuable than wearable technologies.