Abstract
Non-contact vital sign monitoring is essential for medical and rescue applications where physical contact is impossible or undesirable. This paper introduces a novel signal processing pipeline, MAE (MTI-Autocorrelation-EEMD), designed to extract respiration and heartbeat signals from millimeter-wave radar data. The technique utilizes Moving Target Indicator (MTI) filtering to suppress static clutter, autocorrelation to identify the subject's range, and Ensemble Empirical Mode Decomposition (EEMD) to separate overlapping frequency components. Experimental validation using a 77 GHz radar shows that the MAE algorithm maintains high accuracy and low error rates even in extreme noise conditions (SNR as low as -17dB), outperforming traditional PCA and EMD-based approaches.
Introduction
Traditional vital sign monitoring relies on contact-based sensors like ECG leads or pulse oximeters. However, these are often unsuitable for patients with severe burns, fragile neonates, or victims trapped in disaster rubble. Non-contact radar systems offer a promising alternative by detecting the minute chest wall displacements (micro-motions) caused by breathing and cardiac activity. The primary challenge lies in the extremely low signal-to-noise ratio (SNR) of these signals, which are often obscured by environmental clutter and electronic noise. Existing methods like Principal Component Analysis (PCA) often fail in these environments due to a lack of robustness or "mode mixing," where breathing and heartbeat signals become indistinguishable.
Methodology: The MAE Algorithm
The proposed MAE algorithm addresses these challenges through a three-stage pipeline:
- Moving Target Indicator (MTI): To isolate the human subject from stationary surroundings, MTI filtering is applied. This step effectively cancels out reflections from static objects like walls and furniture, focusing the system on the moving chest wall.
- Autocorrelation Detection: Identifying the correct "range bin" (the distance to the subject) is critical. Since life signals are periodic while noise is stochastic, autocorrelation is used to pinpoint the rhythmic micro-motions of the human target with high precision.
- Ensemble Empirical Mode Decomposition (EEMD): Standard decomposition methods often suffer from mode mixing. EEMD overcomes this by adding controlled white noise to the signal before decomposition into Intrinsic Mode Functions (IMFs). This ensures a clean separation of the high-amplitude, low-frequency breathing signal from the low-amplitude, high-frequency heartbeat signal.
Experimental Setup
The researchers validated the MAE algorithm using real-world data collected with a 77 GHz millimeter-wave radar. A healthy male volunteer was positioned 1.2 meters from the radar. To establish a ground truth for heartbeat frequency, a standard finger-clip heart rate monitor was used simultaneously. To test the robustness of the algorithm, varying levels of noise were manually added to the raw radar data, simulating environments with SNRs ranging from 20dB down to -18dB.
Key Results and Discussion
The MAE technique demonstrated exceptional performance in high-noise scenarios. While traditional methods like PCA completely lost the signal at low SNRs, the MAE algorithm successfully identified vital signs at an SNR as low as -17dB.
In terms of accuracy, the estimated heartbeat frequency (e.g., 1.34 Hz) closely matched the reference ground truth (1.40 Hz). Furthermore, the error rate for the MAE method remained below 30% even in the noisiest conditions tested, whereas competing methods saw error rates escalate toward 80-100% as noise increased. This robustness is attributed to the combination of direct MTI cancellation and the adaptive reconstruction capabilities of EEMD.
Conclusion and Future Work
The MAE algorithm provides a robust framework for non-contact vital sign extraction, particularly in challenging, low-SNR environments. While the current 77 GHz hardware has limitations regarding penetration through thick obstacles, the signal processing improvements offered by MAE represent a significant step toward reliable remote health monitoring and search-and-rescue technology. Future work will focus on improving performance in even more complex environments, such as those involving multiple subjects or thicker barriers.
Summary updated