Radar-APLANC: Unsupervised Radar-based Heartbeat Sensing via Augmented Pseudo-Label and Noise Contrast

TL;DR

Radar-APLANC tackles unsupervised radar-based heartbeat sensing using FMCW radar by leveraging an augmented pseudo-label approach and a Noise-Contrastive Triplet loss to learn heartbeat signals without ground-truth physiological data. It introduces a two-stage learning process: Stage One builds coarse heartbeat representations by contrasting pseudo-labels with noise, and Stage Two refines these labels via an adaptive pseudo-label generator with quality assessment and decision rules. The approach achieves performance close to supervised methods on the Equipleth dataset and a newly collected RHB dataset, while offering improved robustness to noise and better cross-dataset generalization. This work reduces data annotation requirements for radar heartbeat sensing and enhances practicality for real-world monitoring.

Abstract

Frequency Modulated Continuous Wave (FMCW) radars can measure subtle chest wall oscillations to enable non-contact heartbeat sensing. However, traditional radar-based heartbeat sensing methods face performance degradation due to noise. Learning-based radar methods achieve better noise robustness but require costly labeled signals for supervised training. To overcome these limitations, we propose the first unsupervised framework for radar-based heartbeat sensing via Augmented Pseudo-Label and Noise Contrast (Radar-APLANC). We propose to use both the heartbeat range and noise range within the radar range matrix to construct the positive and negative samples, respectively, for improved noise robustness. Our Noise-Contrastive Triplet (NCT) loss only utilizes positive samples, negative samples, and pseudo-label signals generated by the traditional radar method, thereby avoiding dependence on expensive ground-truth physiological signals. We further design a pseudo-label augmentation approach featuring adaptive noise-aware label selection to improve pseudo-label signal quality. Extensive experiments on the Equipleth dataset and our collected radar dataset demonstrate that our unsupervised method achieves performance comparable to state-of-the-art supervised methods. Our code, dataset, and supplementary materials can be accessed from https://github.com/RadarHRSensing/Radar-APLANC.

Figures (4)

• Figure 1: Comparison of radar-based heartbeat sensing methods. (a) Traditional methods simply rely on phase extraction and unwrapping processing, resulting in low signal quality; (b) Existing supervised methods require ground truth signals; (c) Our proposed unsupervised approach generates high-quality predictions without requiring ground truth signals.
• Figure 2: The framework of Stage One. The heartbeat matrix, pseudo-label, and random noise matrix undergo random temporal sampling and power spectrum densities (PSD) transform before being fed into the NCT loss. Within this framework, the PSD of the heart matrix is attracted to that of the pseudo-label while being repelled from the noise PSD.
• Figure 3: The framework of Stage Two with Augmented Pseudo-label Generator. It consists of Quality Measurement Module and Decision-making Module.
• Figure 4: Example heart pulse signals generated by our stage one and stage two models and the ground truth signal.