Reliable Quasi-Static Post-Fall Floor-Occupancy Detection Using Low-Cost Millimetre-Wave Radar

Technical Summary: Reliable Quasi-Static Post-Fall Floor-Occupancy Detection Using Low-Cost Millimetre-Wave Radar

Problem Statement
As the population ages, long-term care (LTC) facilities face increasing pressure to monitor residents for falls while managing staff workloads. While fall-event detection (identifying the moment of impact) is well-studied, the clinical requirement often extends to the "post-fall" state, specifically the "long-lie" scenario where an individual remains on the floor for an hour or longer. This state poses severe health risks, including hypothermia and pressure ulcers. Existing solutions face limitations: wearables can be forgotten or malfunction, and camera-based systems raise privacy concerns and fail under occlusion or poor lighting.

Radio-frequency (RF) sensing, particularly millimetre-wave (mmWave) radar, offers a privacy-preserving alternative robust to lighting and occlusion. However, detecting a post-fall state is distinct from detecting a dynamic fall event. Once a fall occurs, the subject becomes quasi-static (minimal motion). This creates a significant technical challenge: the radar return from a stationary human is weak and can be easily masked by ground reflections, multipath effects from furniture, and background clutter. Current literature lacks robust solutions for reliably detecting this specific quasi-static floor occupancy in realistic, furnished environments.

Methodology
The study utilizes an off-the-shelf 60 GHz Frequency-Modulated Continuous-Wave (FMCW) radar (Infineon XENSIV™ BGT60TR13C) to address this challenge.

• Experimental Setup: Data was collected in a furnished room simulating an LTC environment, featuring common clutter (sofas, armchairs, shelves) to induce realistic multipath and partial occlusions. Two radar viewpoints (TV-side and Window-side) were mounted on walls at approximately 2.5 meters. Seven participants were recorded lying in various quasi-static postures (supine, side-lying, knees-bent) across five distinct floor locations (P1–P5) for two minutes each, generating approximately 43,000 frames.
• Signal Processing Pipeline:
  1. Front-end: Standard FMCW processing (Range-FFT and Doppler-FFT) generates Range-Doppler Maps (RDMs). A Moving Target Indicator (MTI) filter is applied to suppress static clutter, though the complex phase is preserved for spatial processing.
  2. Spatial Processing (Comparison): The study compares two approaches for generating Range-Azimuth (RA) maps:
     • Baseline: A vendor-provided Digital Beamforming (DBF) pipeline (Bartlett/Delay-and-Sum), which sums antenna channels with steering weights.
     • Proposed: A preprocessing pipeline based on Minimum Variance Distortionless Response (MVDR)/Capon beamforming. This method aggregates Doppler snapshots near zero velocity to form a spatial covariance matrix. It then solves an optimization problem to minimize output power while maintaining a distortionless response in the target direction. This approach is designed to better suppress interference and enhance weak, quasi-static targets in cluttered environments.
  3. Detection: Both pipelines feed into a 2D Cell-Averaging Constant False Alarm Rate (CA-CFAR) detector. The system treats each frame as a binary decision (occupancy present vs. empty). A False Positive Rate (FPR) cap of 0.1 is enforced during tuning, and a temporal integration rule (e.g., requiring detection in a high fraction of frames over 20 seconds) is proposed for final alarm triggering to mitigate single-frame errors.

Key Contributions

1. Problem Formulation: The authors explicitly frame quasi-static post-fall floor occupancy as a reliability-critical sensing problem distinct from dynamic fall-event classification, setting up experiments under realistic LTC conditions with furniture and occlusions.
2. Algorithmic Innovation: The paper proposes and evaluates an MVDR/Capon beamforming preprocessing pipeline tailored for low-mobility targets. It demonstrates that this method outperforms the standard vendor DBF baseline when paired with a CA-CFAR detector.
3. Comprehensive Evaluation: The study provides a rigorous evaluation across seven participants, multiple viewpoints, and varied floor locations, analyzing performance via mean frame-positive rates, worst-case scenarios, and reliability curves across different success thresholds.

Results
The proposed MVDR-based method demonstrated superior performance over the vendor DBF baseline:

• Overall Accuracy: The mean frame-positive rate improved from 0.823 (DBF) to 0.916 (Proposed).
• Robustness: The proposed method improved or matched the DBF baseline in 94.3% of individual trials.
• Worst-Case Performance: Significant gains were observed in difficult scenarios where DBF struggled (e.g., Subject 1 at TV-side position P1 improved from 0.35 to 0.57; position P5 improved from 0.63 to 0.94).
• Reliability: Reliability curves showed that the proposed method maintains higher coverage even as the success threshold becomes stricter, indicating better performance in the worst-case scenarios critical for clinical safety.
• False Alarms: The proposed pipeline achieved a 100% True Negative Rate (0% FPR) on empty-room data across all subjects, compared to 99% for the DBF baseline.

Significance and Claims
The paper claims that reliable post-fall detection is more challenging than fall-event detection due to the dominance of floor reflections and the lack of motion signatures. The authors assert that their proposed MVDR/Capon preprocessing pipeline effectively mitigates these issues, providing more stable detections and higher reliability across diverse subjects and room layouts.

The study concludes that while the system improves mean performance, its primary value lies in reducing missed detections in difficult layouts and positions (worst-case reliability), which is crucial for safety-critical monitoring. The authors note that the system is not intended to trigger alarms on single frames but rather relies on temporal integration to generate reliable alerts. Future work is identified as expanding the pipeline to a complete workflow combining fall and post-fall detection, as well as integrating post-fall vital sign monitoring.