Your Wi-Fi may already know where you are standing in the room

The quiet number in every Wi-Fi connection

Every Wi-Fi connection carries a small number that most people never see. It is called RSSI, and it reports how strong the signal feels to the device at this moment. A phone, a smart speaker, or an Android TV can read that value through ordinary software, so it sits in plain sight inside many household apps. The number is coarse. It does not describe each tiny piece of the radio wave; it gives one overall strength reading for the link between sender and receiver.

That simplicity is what makes RSSI interesting. More detailed Wi-Fi sensing usually relies on CSI, which tracks amplitude and phase across many subcarriers in a signal. Those measurements can reveal how a body bends or blocks the air around it, but they often need special drivers, elevated permissions, or dedicated hardware. Ordinary consumer devices rarely expose them. A TV may happily report RSSI while keeping the deeper radio details hidden from normal apps.

A test at the University of Iceland showed that this quiet number is not harmless background noise. When shaped into a form close to richer Wi-Fi data, it can feed a model that estimates where a person stands in a room. The result points to a privacy problem: people notice cameras and microphones, but they rarely watch for wireless readings used as clues about movement or presence. A number that seems too weak to matter can become an invisible monitoring channel if software turns it into inference.

A cheap room rig and a model borrowed from richer data

A University of Iceland test showed that ordinary Wi-Fi strength readings can hint at where a person stands in a room. The idea was to reuse an existing model trained on richer Wi-Fi data, not build a new one for cheap boards. Many smart devices expose simple signal strength values through normal software, while detailed subcarrier data often stays hidden. That gap made the borrowed model attractive, because it already knew how to turn Wi-Fi changes into body position.

The original model had learned from detailed per-subcarrier data and depth-camera labels in a similar room. Here it was used for presence, movement, and rough location rather than fine pose. The test space was a cleared campus area about 3.5 by 4 meters. Three low-cost ESP32-C3 boards served as single-channel Wi-Fi transmitters, and one Raspberry Pi 4 received the signal.

Three paths were needed because each transmitter-to-receiver line gives only partial information about a blocking body. Combining changes from three directions allowed a rough estimate of position in depth, not just left or right. Collection began with about two minutes of empty-room signal to set a background baseline. A person then entered and moved while the boards recorded Wi-Fi strength and video captured ground truth at the same time.

Timestamps were synchronized over SSH, and start and stop were semi-automated so people could leave during collection. The raw readings came in dBm, a logarithmic scale that does not match the linear power values expected by the model. So each reading was converted into relative linear power, then scaled and dampened to fit the smaller concrete room. This step made simple strength changes look more like the richer amplitude data used during training.

Hardware quirks also needed care, because cheap boards can introduce uneven noise through cables and power adapters. The team re-recorded data with identical USB-C cables, wall adapters, and surge-protected strips for each transmitter. That reduced a noise difference that had appeared in one board because of its cable and adapter combination. The receiver’s Wi-Fi supplicant was disabled so multiple services would not compete for the same adapter.

The laptop’s own Wi-Fi was also turned off to avoid probe requests that could disturb the readings. The result was a modest rig, built from parts common in hobby projects and small smart devices. It showed that a model trained on richer data can be pointed at cheaper readings without rebuilding the whole sensing stack. That is the practical appeal: ordinary devices may already carry enough signal to support rough human location estimates.

Cleaning the Signal

An empty room is never silent to a Wi-Fi receiver. Walls, power supplies, nearby devices, and the small electronics in the test space all leave faint traces on signal strength. If those traces are left in place, a model can mistake ordinary background wobble for a body moving through the air. The first cleaning step was therefore to record about two minutes of RSSI before anyone entered, so that the quiet state of the room could be subtracted from later readings.

Raw RSSI values are logarithmic, which makes them awkward for a model that expects amplitude-like inputs. Converting dBm readings into relative linear power changed the shape of the data and made small changes easier to compare across time. The subtraction was not a one-size-fits-all operation. Each transmitter had its own quiet-room pattern, so using each device’s own baseline kept the comparison local. A shared average could hide differences between devices and let one noisy path dominate the cleaned signal.

Most tested ways of removing background noise made the empty room look more active than a real person did. That inversion is dangerous because an alarm or tracking system would react to nothing. When only a global mean was removed, the cleaned signal still carried enough structure for the model to reach peak confidence near 0.79. In an empty room, false peaks around 0.65 appeared. Fully normalized data lowered the occupied peak to about 0.61, which made real presence less clear.

The method that kept empty-room confidence below occupied-room confidence was to compare every reading with the quiet baseline from the same transmitter. That choice treated each path as its own instrument rather than forcing all paths into one shared scale. A small noise gate then ignored changes below about 0.001, so faint wiggles did not become fake motion. The remaining signal was smoothed with a sigmoid curve and mapped onto a circular range from -π to +π, making it resemble the phase-like inputs expected by the reused model.

This cleaning work is what turns a noisy strength reading into something the borrowed model can interpret as movement instead of static. Without it, an unoccupied room can trigger the same kind of response that a person produces. With it, the system learns to ignore the room’s own hum and focus on changes that are large enough, consistent enough, and local enough to suggest a body in the space.

Confidence, movement, and the door that opens like a body

When the model was asked to place a person, it did not return one clean answer. For every frame it produced 100 possible body poses, and each pose carried 14 joint locations plus a confidence score. The highest-scoring pose became the main guess, while the other 99 acted as a check on itself. If those alternatives stayed close together, the model was internally consistent, even when its absolute position might still be off.

Movement left a pattern that matched body mass rather than limb detail. The torso joints showed the largest spread among the 100 predictions, which fits the physics of Wi-Fi: a moving trunk disturbs signal paths more strongly than fingers or ankles do. That means the system was reading enough change in obstruction and reflection to say that something large had shifted.

Confidence changed when the raw changes were shaped into an angular form closer to phase data. Without that step, peak confidence reached about 0.72 during occupancy. With it, peak confidence rose to about 0.92, and average confidence stayed around 80 percent while a person was in the room. The shift is important because it shows how quickly a simple signal can look like a reliable presence estimate.

The door added another layer. Opening it produced a distinct change that could be read as an object entering the space, or as a reflectance shift caused by changing distance between one transmitter and the receiver. A doorway moving through its arc can behave like a body: it alters paths, creates a new pattern, and raises certainty even though no person has crossed the threshold.

The Invisible Witness in Every Router

Picture a cleared room where the only witness is a faint wobble in Wi-Fi strength. A person can leave a trace that is harder to notice than a camera or microphone. Ordinary signal-strength readings, the kind many smart devices already collect, can be turned into rough estimates of where someone is and how they are moving. In one controlled test, a model trained on much richer channel data reached about eighty percent confidence while a person moved through a small room. That number is promising, but it does not prove that a living room can be mapped with meter-level accuracy. The system saw broad motion, presence, and changes in posture, yet it also produced confident guesses when the room was empty.

This gap between promise and reliability matters because the hardware is already everywhere. Smart televisions, routers, phones, and other connected objects could become ambient monitors without anyone installing a new camera. That opens useful possibilities for occupancy sensing and privacy-conscious monitoring in homes that value less visible tracking than video. It also creates a harder problem: invisible signals can be read by software that users never think about, so the burden shifts to designers who must build limits, audits, and clear consent into devices from the start.

The real question is not whether Wi-Fi can sense people. The harder question is whether society will decide, before those sensors become normal, what they are allowed to know.