Make the invisible measurable.
The starting question was practical: how much can affordable radio hardware tell us about a space we cannot directly see? Answering it required building the measurement chain—from identified sensors and trustworthy recordings to mathematical models that expose what the measurements can actually resolve.
From the idea
to the implementation.
- 01
Turn the idea into an experiment plan
I broke the goal into acquisition, ranging, signal-change analysis, reconstruction, and material-classification questions. Each stage has an observable result and a reference it must be checked against.
- 02
Connect affordable hardware to a local instrument
I directed NanoC6 firmware and host-side integration, including identified devices, bounded capture commands, integrity-checked serial records, timestamps, and preserved raw recordings. The desktop interface brings experiment controls and results together.
- 03
Implement the reconstruction pipeline
The analysis compares accepted measurements with a baseline, relates each radio link to a spatial grid, and estimates an attenuation-change field with regularization. Coverage and matrix-rank diagnostics show where the available observations are insufficient.
- 04
Make the results reproducible
Recorded sessions can be checked and replayed through alternative analyses. Material experiments require verified reference labels and held-out runs; the wider software lab keeps public-data results separate from local-hardware evidence.
Decisions that
shape the system.
Measure the channel in detail
Channel State Information (CSI) describes the amplitude and phase of a received Wi-Fi signal across frequency bins. The acquisition design preserves those raw values, transmitter and receiver identities, radio settings, and timing so a change can be investigated in context.
Keep ranging and signal contrast separate
Fine Timing Measurement (FTM) supplies native radio-ranging observations. CSI and received-power changes describe a different aspect of the environment. Each has its own calibration and quality checks before contributing to a shared spatial model.
Make reconstruction explain its assumptions
The tomography implementation models how radio links intersect a spatial grid and solves a regularized inverse problem. It reports link coverage, unresolved dimensions, and the influence of smoothing assumptions alongside the estimated field.
Make a wall label earn its meaning
The material-classification code compares matched reference and cross-wall measurements using attenuation, phase, coherence, and temporal stability. Training requires independently verified material labels, measured viewing geometry, and separate validation runs.
- Identify sensors
- Record raw signals
- Compare baselines
- Estimate spatial change
- Validate against references
Inside the implementation
The local instrument combines ESP-IDF firmware, a Python acquisition service, preserved Parquet recordings, and a React/Electron desktop interface. A replay runner checks recording hashes and event order before analysis, allowing the same experiment to be revisited without collecting it again. Camera and photo analysis contribute image-plane observations; they enter metric geometry only when calibration supports that conversion. A separate software-comparison lab evaluates public Wi-Fi sensing implementations and records their hardware and dataset requirements.
INSIDE RADIO X-RAY
A measurement chain.
A testable hypothesis.
The research asks whether changes across multiple Wi-Fi paths can support useful spatial estimates. The implementation makes each transformation inspectable, from radio samples to the proposed map.
Keep the signal’s context.
Raw channel samples retain their signed values and frequency-bin layout. Device identity, timing, radio settings, and drop counters help distinguish an environmental change from a recording problem.
Acquisition + protocol modules
Look for change against a reference.
The wall-analysis module extracts relative attenuation, phase differences, frequency trends, coherence, and temporal stability from compatible reference and cross-wall samples. Antenna orientation and multipath remain alternative explanations.
Wall feature extraction
Solve the inverse problem.
The tomography module relates radio paths to a 2D or 3D grid and fits a smooth attenuation-change field. It reports unobserved cells, matrix rank, and how strongly the reconstruction depends on its assumptions.
Tomography + observability analysis
Compare the estimate with reality.
Material classification requires independently verified examples and separate test runs. Calibrated geometry and reference measurements are needed before an estimated field can support a claim about a wall’s contents.
Reference labels + held-out evaluation
Measured signal changes ≈ path sensitivity × spatial change
The solver adds a smoothness constraint to stabilize this underdetermined problem. A finer display grid does not create finer measured resolution.Research foundations: Espressif ESP-CSI ↗ · Wilson & Patwari · Radio Tomographic Imaging with Wireless Networks ↗. AutiTech’s contribution is the local instrument, integration, experimental workflow, and evaluation—not authorship of those upstream methods.
What the work
demonstrates.
- Verified in the recorded firmware rollout
- 6 boards
- Python and frontend checks in the September 18 run
- 248 + 17
- Native FTM observations retained for analysis
- 52 reports
The work produced a functioning local research application, a recorded firmware rollout to six verified NanoC6 boards, experiment orchestration, raw-data recording, replay analysis, and spatial visualization. A September 18 software verification recorded 248 passing Python tests and 17 frontend tests, plus successful lint, type, and desktop-build checks. A retained FTM experiment produced 52 reports, including 49 native successes; these establish acquired ranging observations, with distance accuracy still to be calibrated.
Why it matters to a client
This is the kind of work needed to make an ambitious sensing idea testable: connect embedded devices to usable software, preserve the measurements, implement the mathematics, and define experiments that can distinguish a useful signal from a convincing-looking picture.
Known locally as Radio Scans. “Radio X-Ray” is the project name; the instrument uses radio signals, not ionizing X-rays. Work is confined to authorized spaces and local processing.