Description
Detecting that a person has fallen — typically in a domestic or care setting — from passive wireless observation. Falls are short, high-energy events with a characteristic Doppler signature on WiFi CSI and mmWave radar. The clinical motivation is high: unattended falls in elderly populations have severe outcomes, and a non-wearable detector avoids the abandonment problem of wearable PERS devices. Fall detection is the highest-impact safety application of wireless human sensing.
Why it's hard
- Class imbalance: real falls are rare in any training corpus and synthetic falls are not the same.
- False positives are operationally costly (loss of caregiver trust).
- Distinguishing a fall from sitting-down-quickly or dropping-an-object requires temporal context.
- Per-home calibration drift is real — furniture rearrangement breaks trained models.
- Multi-person homes complicate per-individual fall attribution.
Common approaches
- CSI Doppler / spectrogram CNNs trained on simulated + acted falls.
- Anomaly-detection framing on top of activity-recognition models.
- Multi-modal fusion (CSI + radar + audio) for robust detectors.
- Continual learning for per-home adaptation.
Source Papers
- ahmad2024_8639 ↗ — WiFi-based human sensing with deep learning (fall detection covered).
- ullmann2023_0ac3 ↗ — radar-based continuous HAR (fall a primary class).
- guarino2026_e72c ↗ — CSI-based WiFi sensing datasets and reproducibility.
- chen2023_5cbd ↗ — cross-domain WiFi sensing (fall under domain shift).