A Comparative Study of Methods of Person Identification Using Radar-Extracted Heartbeat Signals

In recent years, non-contact biometric identification has garnered significant attention owing to its flexibility and its capacity to ensure privacy and confidentiality. Previous research has predominantly centered on harnessing cardiac radar signals detected by radar systems. This paper aims to compare two methods for person identification utilizing heartbeat signals extracted from radar data. In the first approach, we directly employ the heartbeat time series data as input for a deep learning model called InceptionTime dedicated to person identification. The second method explores the viability of using spectrograms, which are images generated from cardiac radar heartbeat signals, combined with other deep learning models, ResNet and a small convolutional neural network (CNN) model built by us, to achieve the same goal. Additionally, we assessed the robustness of both methods by introducing white Gaussian noise into the signals and evaluating their performance under different noise levels. For the second method, we implemented three image denoising methods: autoencoder, deep image prior (DIP), and Dilated-Residual U-Net (DRU-Net) to denoise the spectrograms after adding noise in the time domain. Experimental results reveal that the accuracy of person identification using spectrograms is 100% without adding noise based on the public dataset. The identification accuracy is the same as that using time series data. Following the introduction of noise, the accuracy of both methods decreases; however, the method based on spectrograms exhibits a more pronounced decline in accuracy compared to the approach using time series data directly especially for the small CNN model we built. The image-based method with DIP image denoiser shows outstanding performance when the value of noise is high compared with InceptionTime.