Doppler-Domain Respiratory Amplification for Semi-Static Human Occupancy Detection Using Low-Resolution SIMO FMCW Radar

Technical Summary: Respiratory-Amplification Semi-Static Occupancy (RASSO) for Low-Resolution SIMO FMCW Radar

Problem Statement
Radar-based sensing offers a privacy-preserving alternative to cameras for monitoring human occupancy in settings like long-term care. However, detecting quasi-static presence (lying, sitting, or standing with only subtle micro-motions) remains a significant challenge for low-resolution Single-Input Multiple-Output (SIMO) Frequency-Modulated Continuous-Wave (FMCW) radars. In these scenarios, the micro-Doppler signatures generated by respiration and minor postural adjustments are concentrated near zero velocity. Consequently, this energy is often buried under static clutter (walls, furniture) and the DC component of the radar signal. Standard processing pipelines, which utilize uniform Doppler grids, lack the resolution to distinguish these faint signals from background noise, leading to poor signal-to-noise ratios (SNR) and unreliable detection.

Methodology
The authors propose RASSO (Respiratory-Amplification Semi-Static Occupancy), a signal processing technique designed to enhance the detection of semi-static targets before spatial beamforming. The core of the methodology involves an invertible, non-linear warping and resampling of the Doppler (slow-time FFT) axis.

1. Doppler-Domain Warping: RASSO applies a logarithmic-like mapping to the Doppler frequency axis. This transformation densifies the sampling grid around near-zero frequencies (where quasi-static energy resides) while compressing the sampling at higher frequencies. The mapping is defined as $D = \text{sgn}(f) \frac{f_e}{\log 2} \log(1 + \frac{|f|}{f_e})$, where $f_e$ is an edge frequency parameter (empirically tuned to 0.03 Hz).
2. Phase Preservation: Unlike methods that accumulate magnitude heatmaps (which lose phase information), RASSO performs complex linear interpolation on the real and imaginary parts of the Range-Doppler (RD) data. This preserves the per-antenna phase relationships required for subsequent digital beamforming.
3. Processing Pipeline: The warped RD data is fed into a Capon (Minimum Variance Distortionless Response, MVDR) beamformer to generate Range-Azimuth (RA) maps. These maps are subsequently processed using Cell-Averaging Constant False Alarm Rate (CA-CFAR) detection.
4. Evaluation Framework: The study utilizes a real-world dataset collected in a nursing home using an Infineon BGT60TR13C (1Tx-3Rx) radar. The evaluation compares RASSO-enhanced pipelines against:
   • Standard Digital Beamforming (DBF) from the vendor SDK.
   • A conventional Capon-based baseline without warping.
   • Recent state-of-the-art baselines (E-RESPD/MCROOD).
   • Both knowledge-driven (CFAR) and data-driven (CNN, CNN-LSTM) classification approaches.

Key Contributions

• Algorithmic Innovation: The introduction of RASSO, an invertible Doppler-domain warping technique that increases the effective resolution of near-zero micro-motions without discarding phase information.
• Empirical Validation in Realistic Environments: Unlike many studies relying on controlled laboratory data, this work validates the method on a dataset collected in a long-term care facility, featuring realistic multipath, occlusions, and natural micro-movements.
• Dual-Approach Evaluation: The paper demonstrates that RASSO improves performance for both traditional knowledge-driven detection (CFAR) and modern data-driven models (Deep Learning), showing that the preprocessing benefits the learning process itself.

Results

• Signal Quality: RASSO-enhanced RA maps exhibit sharper target peaks, lower background variance, and higher effective SNR. The average SNR improved from 6.88 dB (conventional) to 9.55 dB (RASSO).
• Knowledge-Driven Detection (CFAR):
  • RASSO-RA achieved an Area Under the Curve (AUC) of 0.981, outperforming the conventional Capon baseline (0.920) and the E-RESPD baselines.
  • At a False Alarm Rate (FAR) of 1%, RASSO achieved a recall of 0.920, compared to 0.643 for the conventional baseline.
  • In cross-subject generalization tests, RASSO improved the Macro-F1 score by approximately 6–10 percentage points over the non-warped baseline.
• Data-Driven Classification:
  • A frame-based SimpleCNN achieved 95–99% accuracy with RASSO inputs.
  • A sequence-based CNN-LSTM achieved 99.4–99.6% accuracy.
  • Bootstrap uncertainty quantification confirmed statistically significant gains (95% confidence intervals strictly above zero) in Macro-F1 scores, with improvements ranging from 2.6 to 3.6 percentage points over the conventional pipeline.

Significance and Claims
The paper claims that simple Doppler-domain warping prior to spatial processing can materially improve semi-static occupancy detection with low-resolution radar in real clinical environments. The authors emphasize that RASSO addresses the specific limitation of low-resolution radars struggling with near-zero Doppler energy by reallocating sampling density to the region of interest. The results suggest that this preprocessing step not only aids traditional threshold-based detection but also provides more discriminative inputs for lightweight neural networks, enabling robust, privacy-preserving monitoring of elderly individuals in long-term care settings without the need for high-resolution hardware or invasive wearables. The study concludes that this approach supports dependable deployment beyond controlled laboratory settings.