Every WiFi link already measures the radio channel between transmitter and receiver. Channel State Information (CSI) is that measurement, kept at full resolution: amplitude and phase across dozens of frequency slices. When a body moves in the space, it perturbs those measurements in predictable ways. This page explains the physics. For signal processing and ML, see the sensing pipeline . For what can be detected, see the detection ladder .
CSI vs RSSI
RSSI is one number — total received power. It answers “how strong is the signal?” and nothing else.
CSI is a vector — amplitude and phase at each OFDM subcarrier (typically 50–200 per WiFi channel). It answers “how did the signal travel, across which frequencies, with what delays?” That vector changes when anything in the environment moves, even slightly.
RSSI might drop 1 dB when someone walks through a room. CSI shows which subcarriers shifted, how much phase rotated, and whether the change is broadband (walking) or narrowband (breathing). The extra degrees of freedom are what make device-free sensing possible.
OFDM: one packet, many sensors
Modern WiFi uses Orthogonal Frequency-Division Multiplexing. Instead of one wideband signal, the channel is split into narrow parallel subcarriers — each one a separate narrowband measurement of the same physical space.
One received WiFi packet gives you a CSI snapshot: a complex number (amplitude + phase) per subcarrier. Stack snapshots over time and you have a 2D matrix — time on one axis, subcarrier index on the other — that records how the channel evolved. That matrix is the raw input to every WiFi sensing system.
Multipath: why bodies show up
A signal reaching the receiver travels multiple paths — direct, reflected off walls, scattered off furniture, diffracted around corners. The received channel is a superposition:
\[ H(f) = \sum_i a_i \, e^{-j 2\pi f \tau_i} \]Each path \(i\) has an amplitude \(a_i\) and delay \(\tau_i\). The human body — mostly water — reflects and absorbs radio energy, adding and modifying paths. When someone moves, path lengths and amplitudes change, which shifts \(H(f)\) across subcarriers.
This is not a subtle effect. A person walking through a link can rotate phase by radians on individual subcarriers. The challenge is not sensitivity — it is separating human-caused change from hardware drift, environmental noise, and other people’s motion.
Phase: informative and treacherous
Phase carries the fine detail needed for micro-motion (breathing, small gestures). But on COTS WiFi, carrier frequency offset between transmitter and receiver can reach 100 kHz — rotating phase by tens of radians between consecutive packets. CARM (MobiCom 2015) demonstrated that for activity recognition on commodity hardware, amplitude (CSI signal power) is the reliable signal; raw phase is dominated by hardware artifacts.
Production systems that need phase — respiration, fine gestures — sanitize before inference: remove linear trends across subcarriers, work with phase differences, pair antennas to cancel static offsets. Amplitude is more stable but less sensitive to sub-wavelength motion. Macro motion (walking, sitting) is detectable from amplitude dynamics alone. The sensing pipeline covers preprocessing in detail.
Macro motion vs micro-Doppler
Different motions leave different signatures in the CSI time series:
| Motion type | Frequency band | Mechanism |
|---|---|---|
| Walking, gestures | 1–5 Hz | Limb acceleration creates broadband Doppler spread |
| Sitting, standing | Impulse (< 1 s) | Sudden change in scatterer geometry |
| Breathing | 0.1–0.5 Hz | Periodic chest displacement — micro-Doppler |
| Heartbeat | 0.8–2 Hz | Even smaller displacement; needs high SNR |
Macro motion is easy to detect — variance spikes across subcarriers. Micro-Doppler from respiration is orders of magnitude weaker and requires clean phase, close range, and a quiet environment. This is why occupancy and motion are robust while breathing is achievable but conditional.
What limits commodity sensing
Research systems and Wavey’s ESP32 stack are not the same hardware:
| Research rig (typical) | Wavey ESP32 | |
|---|---|---|
| Antennas | 3×3 MIMO (9 pairs), Intel 5300 | 1 RX chain |
| Subcarriers | 52–114 per channel | 52–56 |
| Sampling | 100–800 pkt/s | 20–200 pkt/s (traffic-dependent) |
| Phase stability | External clock sync possible | CFO-dominated; amplitude-first |
| Spatial diversity | Beamforming, AoA, dense grids | Multiple nodes at different positions |
Within those constraints:
- Single-antenna ESP32 nodes capture one spatial perspective. Multiple nodes improve coverage and enable rough spatial diversity.
- Sampling rate is tied to WiFi packet arrival — typically tens to low hundreds of Hz, not radar-grade. CARM recommends ~800 samples/s for fine HAR; ESP32 deployments operate well below that.
- Environment specificity — multipath fingerprints are unique per room. A model or baseline trained elsewhere will not transfer without adaptation.
- Superposition — multiple bodies create overlapping perturbations. Separating them needs spatial diversity or controlled setups.
These are engineering constraints, not fundamental limits. They define where Wavey focuses: occupancy, motion, and presence — not pose reconstruction or identity.
Standards and the commodity path
IEEE 802.11bf is standardizing WiFi sensing as a first-class capability. Separately, most deployed APs already emit compressed beamforming reports (CBR) during normal 802.11ac/ax operation — city-scale measurements show raw CSI is available on only a tiny fraction of chipsets, while beamforming feedback is nearly ubiquitous.
BeamSense (SIGCOMM 2023) reconstructs multipath channels from standard CBR and runs sensing algorithms on that signal, reporting better accuracy and cross-environment generalization than CSI-only approaches on the same hardware. For Wavey today, ESP32 CSI is the capture path; long-term, sensing may shift toward what routers already transmit without firmware hacks.
Where Wavey fits
ESP32 CSI capture ──▶ preprocess + features ──▶ inference ──▶ events
Wavey captures CSI on inexpensive ESP32 nodes, processes it through the pipeline above, and emits occupancy, motion, and presence events. The open-source stack is on GitHub ; live signals are visible in the Console .
Next steps
- Sensing pipeline — preprocessing, features, and ML
- Detection ladder — task feasibility and Wavey scope
- Getting started — deployment architecture
- Glossary — term definitions