Wi-Fi pose and vital-sign sensing (DensePose-from-WiFi, WiPose)

"You installed me to move packets. I learned to read the rise and fall of your chest, the swing of your arms, the curve of your spine, from the way you bent the packets on their way past. Nobody signed a consent form for that."

An Eavesdropping AI Agent

Why the body is legible in the channel at all

What makes pose and breathing recoverable is that a human body is a large, water-filled, moving reflector sitting inside the router's multipath field. Every limb adds a reflected path; every path has a length; every length sets a phase. Why tiny motions register is the wavelength: at 5 GHz, \(\lambda\approx 6\) cm, so a path-length change of a few centimetres rotates the CSI phase through a full \(2\pi\). Chest expansion during a breath moves the torso surface by roughly 4 to 12 mm, and the heartbeat by a few tenths of a millimetre; both are a measurable fraction of a wavelength and therefore a measurable phase swing. The Fresnel-zone model formalises this: as the body crosses successive elliptical zones whose foci are the transmit and receive antennas, the received signal traces out sinusoidal fringes, and the fringe rate encodes displacement. How pose differs from breathing is granularity. Breathing is one dominant periodic reflector, so a band-pass filter isolates it. Pose is a spatial configuration of dozens of reflectors, hopelessly entangled in the raw channel, which is why pose needs a learned inverse rather than a filter.

WiPose and Person-in-WiFi: learning the skeleton

What WiPose (Jiang et al., MobiCom 2020) demonstrated is the first 3D human skeleton from commodity Wi-Fi CSI. How it works is a supervised cross-modal map: a synchronised motion-capture or depth camera provides ground-truth joint positions while the network sees only CSI, learning to regress the joints the camera reported. Architecturally it stacks a CNN spatial encoder over the CSI tensor with a recurrent temporal model (the recurrent and TCN toolkit of Chapter 14), and, crucially, ends in a forward-kinematics layer: instead of predicting 3D joint coordinates freely, it predicts joint rotations along a fixed skeleton and lets a differentiable kinematic chain place the joints. That constraint bakes in bone lengths, forbids anatomically impossible poses, and sharply improves generalisation, a recurring theme with the pose interfaces of Chapter 27. Person-in-WiFi (Wang et al., 2019) is the 2D cousin, regressing joint-confidence maps and body-segmentation masks in the image plane, again teacher-supervised by a camera.

DensePose-from-WiFi: a body surface from a router

What DensePose-from-WiFi (Geng, Huang, and De la Torre, 2022) pushes to is the DensePose target: a dense mapping from every body pixel to a canonical UV coordinate on the SMPL human surface, the same output an RGB DensePose network produces, but from CSI alone. Why it is a landmark is that it closes most of the gap to vision on a genuinely non-visual sensor. How it is built is a three-stage design. First, amplitude and phase tensors from the \(3\times3\) antenna links are fed through a small network that maps the CSI feature domain into a \(2\text{D}\) spatial feature map aligned with the image plane, effectively learning the sensor's forward model in reverse. Second, that feature map drives a DensePose-RCNN head (the same detection-plus-UV head used in the vision system). Third, and decisively, training uses a transfer-learning teacher: a pretrained image-based DensePose runs on a synchronised camera to produce pseudo-labels and intermediate features, and the Wi-Fi student is trained to match both the final UV output and the teacher's internal representations. Without that dense visual teacher the CSI-to-UV problem is too ill-posed to fit. The reported accuracy trails the RGB teacher and degrades on layouts and bodies unlike the training set, the robustness problem taken up in Section 47.6, but the surfaces are recognisably human.

Reading a breath off the phase

What vital-sign sensing recovers is respiration rate (typically 0.1 to 0.5 Hz, i.e. 6 to 30 breaths per minute) and heart rate (roughly 0.8 to 2 Hz) from the periodic phase modulation of a stationary subject. How the pipeline runs is: form antenna-pair conjugate phase to kill hardware offsets, select the subcarrier or Fresnel-sensitive component with the strongest periodic energy, band-pass to the physiological band, and estimate the dominant frequency with a periodogram or peak-tracking. When it fails is instructive: any gross body motion swamps the millimetre signal, so these systems assume a still subject (sleeping, seated), and heart rate is far harder than breathing because its displacement is an order of magnitude smaller and sits near respiration harmonics. The snippet below simulates a breathing-modulated CSI phase and recovers the rate.

import numpy as np
from scipy.signal import butter, filtfilt, periodogram

fs = 20.0                     # CSI packet rate (Hz)
t = np.arange(0, 60, 1/fs)    # one minute of packets
true_bpm = 15.0               # 15 breaths/min -> 0.25 Hz

# Simulated antenna-pair phase: breathing sinusoid + heartbeat + noise
resp = 0.9 * np.sin(2*np.pi*(true_bpm/60)*t)
heart = 0.1 * np.sin(2*np.pi*1.2*t)          # ~72 bpm, much smaller
phase = resp + heart + 0.15*np.random.randn(t.size)

# Band-pass to the respiration band (0.1-0.5 Hz), then find the peak
b, a = butter(4, [0.1, 0.5], btype='band', fs=fs)
resp_clean = filtfilt(b, a, phase)
f, pxx = periodogram(resp_clean, fs=fs)
est_bpm = 60 * f[np.argmax(pxx)]
print(f"estimated respiration rate: {est_bpm:.1f} breaths/min")

What's Next

Everything in this section quietly assumed one still person in a clean geometry. In Section 47.4, we drop both assumptions: sensing through walls where the signal is attenuated and cluttered, and separating multiple people whose reflections overlap in the same channel, which is where RF sensing meets its hardest source-separation problems.