The most useful sensing question is also the easiest to answer reliably: is anyone here? Binary occupancy asks for a single bit — empty or occupied — which tolerates noise far better than counting people or classifying pose.
Change-point detection on a baseline
Wavey learns an empty-room baseline: the mean amplitude profile, typical subcarrier variance, and phase structure when no one is in the space. Live CSI compared against this reference uses change-point detection — when the deviation exceeds a threshold sustained over a window, occupancy is declared.
An alternative feature family from the building-energy literature uses signal tendency index (STI) — shape similarity between adjacent CSI amplitude curves — with greedy subcarrier selection to pick frequencies most responsive to human presence. WiFree reported 99.1% occupancy detection on building-scale deployments. The mechanism differs from a single variance threshold, but the decision is the same: is the current CSI pattern consistent with an empty room?
The signal does not need to be clean. It needs to be different from empty. A person sitting still perturbs multipath enough to shift the baseline; breathing adds a periodic micro-Doppler component that an empty room never produces.
Threshold selection trades sensitivity against false triggers from environmental drift (HVAC cycling, a door opening). Hysteresis — requiring sustained deviation before flipping state, and sustained return before clearing — prevents flickering at the boundary.
Why binary occupancy is the sweet spot
Occupancy is robust because the decision boundary is wide:
- Noise tolerance — you do not need to classify what happened, only that something happened.
- Single-link sufficiency — one TX-RX pair can detect presence in the channel; spatial diversity helps but is not required.
- Low label burden — empty vs occupied is easy to label or self-supervise from known-empty periods.
This is why building energy systems and smart-home HVAC start here. The detection ladder places occupancy at rung 2 — the first task that works reliably on commodity hardware.
Not the same as phone counting
MAC-address analytics count devices, not people. One person with a phone, watch, and laptop registers as three. Five people sharing one laptop register as one. Device identifiers can constitute personal data under GDPR.
CSI occupancy reads how bodies disturb the radio field — no identifiers, no device dependency. Everyone in the space is detected whether or not they carry anything. For analytics semantics, see occupancy analytics .
Building and HVAC control
Occupancy-aware HVAC does not need person count. It needs a reliable signal that the space is in use — including when occupants are still. A conference room with six people sitting quietly should not trigger ventilation shutdown because no one is waving.
CSI occupancy provides that hold signal. Pair it with motion for activity-responsive scenes in smart-home automation .
Further reading
- Baselines and generalization — drift, cross-site transfer
- Sensing pipeline — preprocessing before occupancy inference
- Getting started