The radio waves you're already paying for

The WiFi access point bolted to your office ceiling is already doing something quietly remarkable. Every packet it sends triggers a measurement: how did the radio channel distort that signal across each of its subcarrier frequencies? The result is a vector of complex numbers called Channel State Information (CSI) — computed millions of times a day, used to tune the equaliser, and then thrown away. Wireless sensing is the argument that throwing it away is a mistake. Those discarded snapshots carry a passive, device-free record of every body moving through the room. No camera. No wearable. No opt-in required.

A communication measurement moonlighting as a sensor

CSI is not a sensing technology bolted onto WiFi as an afterthought — it is a measurement WiFi already makes for communication, and wireless sensing is the discipline of reading human presence back out of it.

To understand why that is even possible, you need one piece of geometry. Every transmitter–receiver pair is surrounded by a set of ellipsoidal shells called Fresnel zones — regions of space where reflected signals arrive in-phase or out-of-phase with the direct path. When a human body steps inside the first Fresnel zone, the innermost shell where the signal is most sensitive, it diffracts the wave sharply and produces a measurable drop in amplitude. Step outside that zone and the body acts more like a reflector, adding a phase-shifted copy of the signal. Either way, the occupant's position is encoded in the amplitude and phase of every subcarrier wu2022_75d3.

The practical trick for reading that encoding cleanly is the CSI ratio: dividing the CSI measured on one antenna by the CSI on a second antenna. Hardware imperfections — clock drift, amplifier noise — appear identically on both antennas and cancel in the division, leaving only the room's geometry. The physics is not magic; it is careful arithmetic on a measurement WiFi was already making.

ma2020_4782 formalises this in a canonical survey: on commodity hardware like the Intel 5300 802.11n CSI Tool, a 20 MHz channel yields 30 subcarriers per packet; on the Atheros CSI Tool, 52. Each subcarrier is an independent probe of the channel. Together they form a spatial fingerprint of the room's current state.

More subcarriers mean finer-grained spatial information — and more signal-processing headaches, which later episodes will dig into.

Why this matters beyond the lab

A reasonable objection: PIR sensors (passive infrared detectors that trigger on body heat crossing their field of view) cost a few dollars and work reliably. Cameras cost more but give rich data. Why bother with CSI?

The answer sits at the infrastructure layer. The radio industry is converging on Integrated Sensing and Communication (ISAC) — a design philosophy now entering standards bodies that treats sensing as a first-class service of the radio stack rather than a side effect wang2022_2397. Future WiFi and 5G base stations will dedicate part of their radio resources explicitly to sensing. The CSI-sensing research happening in university labs today is, structurally, a preview of what commodity infrastructure will do by default. That is why this is an IoT question and not just a lab curiosity — and part of what drove the hands-on CSI work we kicked off in meetings/2026-05-13-csi-experiments.md.

The gap that keeps cameras employed

Here is the honest state of play. A CSI-plus-ML pipeline called Wi-CaL can count up to seven people in a small room at roughly 63% accuracy and outperforms passive WiFi radar in line-of-sight conditions choi2022_17c2. That is genuinely useful — and a long way from drop-in camera replacement.

The structural reason is what chen2023_5cbd calls the cross-domain problem. Train a CSI model in one room, with one antenna placement, on one group of people, and it learns the fingerprint of that room. Change the furniture, move the access point a metre, swap in a different device pair, or ask it to recognise someone it has never seen — and accuracy collapses. Environment, antenna geometry, user identity, and motion speed each shift the CSI distribution enough to break the model. This is not a tuning problem; it is a structural one. The room's radio fingerprint is specific to that room in ways that are hard to factor out.

That cross-domain generalisation gap is the load-bearing engineering challenge this series will keep returning to. Every technique we examine — denoising, domain adaptation, graph-based representations — is ultimately aimed at making CSI models that survive contact with the real world.

If a pipeline can count seven people in a controlled lab at 63% accuracy today, what would it concretely take — in labelled data, in antenna placement choices, in algorithm design — to make that accuracy hold when someone moves the access point one metre to the left?