Wi-Fi CSI fall detection: features that work, and the false alarms you must handle

2026-10-11 · 2 min read

Fall detection is the highest-value use case in Wi-Fi sensing (elder care — see WiFall ^[2] and WiCare ^[3]) and the easiest to get wrong. Presence detection answers "is the channel disturbed?"; fall detection must separate a fall from sitting down, putting down a bag, or a door slamming.

1. What a fall actually looks like in CSI

So the discriminative features are burst energy + bandwidth + the post-event quiet period, not raw variance (feature taxonomy: CSI sensing survey ^[1]).

2. A workable feature set

import numpy as np

FS = 95.0                                  # measured records/s (resample by timestamp!)
WIN = int(1.5 * FS)                        # 1.5 s analysis window

def fall_features(H):                      # H: (time, subcarriers) amplitude
    d = np.abs(np.diff(H, axis=0)).mean(axis=1)         # motion envelope
    win = d[-WIN:]
    energy = float(win.max())
    rise = float(win.max() - win[: len(win) // 4].mean())  # abrupt onset
    quiet_after = float(win[-int(0.5 * FS):].mean())       # post-impact stillness
    # Doppler bandwidth via STFT of the phase difference
    z = np.fft.rfft(win - win.mean())
    spread = float(np.sum(np.abs(z) ** 2) / (np.abs(z).max() ** 2 + 1e-9))
    return dict(energy=energy, rise=rise, quiet_after=quiet_after, spread=spread)

Then classify with explicit thresholds (start simple; a fall is rise > R and spread > S and quiet_after < Q) and require confirmation within 5 s.

3. Killing false positives

Add a pre-event context window (walking for 2-5 s before) and require it. That single rule removes most single-burst false alarms without deep learning.

4. Data and evaluation

• record your own: label fall / sit / drop / walk; 30-50 events per class is a workable start;
• report per-class precision/recall and the false-alarms-per-day number — that is what eldercare buyers ask;
• publish the recording conditions (room, distance, height) with the dataset, or nobody can reproduce it.

5. Legal and ethical boundary (do not skip)

Fall detection sits next to medical claims. Keep the wording precise:

• say "detects fall-like events" and "not a medical device";
• never promise emergency response on its own;
• get consent from the person being monitored — this is monitoring, even without a camera.

Related: respiration detection · vs mmWave radar · Home Assistant integration