A Raspberry Pi platform that fuses WiFi channel data, radar and a mesh of radio links into one picture of a room, and tells you honestly which sensor saw what.
HERMIT stands for Heterogeneous Edge RF and Motion Intelligence Toolkit. It is built to notice movement, a person sitting perfectly still, and the rhythm of their breathing, while they carry nothing.
Short-window detection of movement from shifts in the WiFi channel.
Presence of someone not moving, measured against a 20 s calibration of the empty room.
Respiration accepted between 8 and 30 breaths per minute, with split-half validation.
That is the problem HERMIT starts from. Radio can still see a person who is sitting, sleeping or reading, because a breathing body keeps changing the channel it sits in.
A blob, not a silhouette.
Every source publishes onto one frame bus. Detectors turn frames into events, and fusion combines them into a single state that can always explain itself, sensor by sensor.
WiFi channel state, SDR sweeps, radar, environment and GNSS publish typed frames.
Guard carriers removed, gain normalised, outliers filtered, live subcarriers selected.
Motion, stillness and breathing run on separate windows, each with its own threshold.
Bounded log-odds: no single sensor can dominate, and every vote is recorded.
Desktop app, local dashboard and a JSONL event log. Events, not raw channel data.
Buy uncorrelated failure modes, not redundancy.
The design principle behind the sensor mix. WiFi, radar, infrared and radio mesh each fail in different conditions, so fusing them is worth more than doubling any one of them.
Channel state. Captured through nexmon_csi on the Pi’s bcm43455c0 radio. With a single antenna the chain uses amplitude only: guard-subcarrier removal, per-frame gain normalisation, a Hampel filter, live-subcarrier selection, fractional deviation and PCA. Presence compares a 10 s window against the calibrated empty room at an 8-sigma threshold. Heart rate (48-120 bpm) is always reported as low confidence.
Fusion. Evidence combines in bounded log-odds with each source capped at ±2.5, an 8 s evidence half-life, a prior of 0.15 and an alert threshold of 0.75. An explain call lists each sensor’s signed vote, including disagreements.
Radio tomographic imaging. ESP32 nodes beacon over ESP-NOW and report received signal strength. The image solves y = Wx + n with Tikhonov regularisation, and an alpha-beta tracker follows people, also accepting LD2450 radar targets. A mesh of N nodes gives N(N-1)/2 links: 20 nodes make 190.
RF monitoring. SDR sweeps learn a 300 s site baseline, then flag new emitters with an adaptive noise floor, a 16 dB threshold with hysteresis and persistence checks. A dedicated detector watches GNSS L1 for jamming.
A reliable room-occupancy node is planned at about $110. Planning figures in early-2026 US dollars, not quotes.
| Part | Role | Approx. |
|---|---|---|
| Raspberry Pi 4B, 4 GB | Compute and the WiFi radio that measures the channel | $55 |
| HLK-LD2410C | 24 GHz radar presence that still sees a person who is not moving | $5 |
| HLK-LD2450 | 24 GHz radar tracking up to three targets to about 6 m | $8 |
| RTL-SDR Blog V3 or V4 | Receive-only spectrum monitoring up to 1.766 GHz | $30-40 |
| HC-SR501 PIR | Independent motion confirmation | $2 |
| BME680 | Temperature, humidity and VOC corroboration | $10 |
| u-blox NEO-8M with PPS | Position and microsecond timing | $15 |
| ESP32 WROOM-32 | One radio tomography mesh node, all-in | $9 |
Five printed trays on a 150 × 110 mm footprint, 167 mm tall. Most of the rules exist to protect a measurement:
A hand-separable 1:2 presentation model, 75 × 55 × 84 mm on magnets and pins, carries the sensors as loose tiles.
The same standard as the rest of this site. Nothing below is rounded up.
Physics limits apply too. Brick and concrete cost 10-20 dB or more, steel bulkheads stop everything, and one WiFi link at 20 MHz cannot resolve position finer than about 7.5 m. Position needs the mesh.
A sensor that notices people who cannot see it needs a position before it needs a feature list. This is ours.
Would you be comfortable telling the people this sensor can detect exactly what it does?
The test every deployment has to pass. If the answer is no, the answer is no.
HERMIT and the terahertz programme share one discipline: sense what matters about a person, publish exactly what the sensor can and cannot do, and never build an imaging device when a measurement will do.
