Skip to content

wifi-sensing.md

Latest commit

 

History

History
120 lines (97 loc) · 5.66 KB

wifi-sensing.md

File metadata and controls

120 lines (97 loc) · 5.66 KB

English · 中文

Wi-Fi sensing and where diting sits

If you've heard about Wi-Fi being able to "detect motion through walls", "count people in a room", or "monitor breathing without a camera" — yes, that's a real research field. It's called Wi-Fi sensing. Whether diting can do it is a question we get asked often enough to deserve a written answer.

Short answer

No, and on purpose. diting is a macOS terminal monitor for Wi-Fi roaming and signal quality. The fancy sensing capabilities above all require Channel State Information (CSI) — per- subcarrier amplitude and phase data — which macOS does not expose to user-space code. We work with what CoreWLAN gives us: RSSI, channel, BSSID, etc. That's enough to build a useful dashboard; it's not enough to do pose estimation.

The capability ladder

Wi-Fi sensing isn't one thing. It's a spectrum from "feasible today" to "research demo only":

Tier What Hardware needed Reliability
0. RSSI variance "RF environment is stirring" / coarse motion Any Wi-Fi card Decent for line-of-sight motion
1. CSI presence Empty vs occupied, distinguish 1 person walking ESP32, Intel 5300, or research NIC Good with calibration
2. CSI activity recognition Walking / sitting / falling, gestures Above + trained ML model Decent in trained environment
3. CSI vital signs Breathing rate, heart rate Above + clean SNR + bandpass filtering Real research, controlled environments
4. CSI pose / through-wall 17-point body pose, multi-person tracking Multistatic mesh + heavy ML Research only, no production-grade implementation

diting sits firmly in tier 0, and only because RSSI is the data we already collect for the roaming dashboard.

Why CSI is not available on macOS

Apple's Wi-Fi driver and firmware are signed, closed, and accessed only via private frameworks that don't expose CSI. The nexmon_csi firmware- patching project — which is the canonical way to get CSI from Broadcom chips — supports Raspberry Pi and certain Linux laptops but not macOS or Apple Silicon Macs. There is no public CSI API on macOS today. Apple Silicon makes the path harder, not easier.

For real CSI work, the practical options are:

These projects exist; we link to them rather than reimplementing them poorly inside diting.

What you should ignore

Search results on this topic are noisy. Be careful with projects that promise tier-3 or tier-4 capabilities (vital signs, pose, through-wall sensing) as turn-key open-source code. The pattern to recognise:

  • Long technical-sounding pipelines with named stages ("gestalt / sensory / topology / coherence / search / model")
  • Specific accuracy numbers cited without paper / benchmark links
  • Claims that work "with any Wi-Fi router" out of the box
  • Star count not matched by working demos or third-party reproductions

These are red flags for AI-generated boilerplate that doesn't run. The 2025-2026 Hacker News thread on RuView is one extended example. The underlying science is real (CMU's WiFi-DensePose paper, MIT's vital-radio work) — but turning that into a pip install is a research-grade project, not a weekend.

What diting adds at tier 0 (v0.7.0)

Without overclaiming, the RSSI / Tx-rate data we already poll supports — and as of v0.7.0 actually drives — three concrete surfaces:

  • An environment-stability indicator, the new Environment line in the Diagnostics panel. Format: Environment stable σ 1.2 dB / 5s. The label is one of stable / active / quiet (the last only when an opt-in calibration baseline from diting calibrate is loaded). NEVER "N people" or "motion direction" — the wording rule is non-negotiable.
  • Motion event logging — when current 5 s σ exceeds 2.5 × trailing 5-min median σ AND the absolute σ exceeds 3 dB, an rf_stir event is appended to the unified events ring. Confidence is high when the spike shows up on >= 2 co-located APs at once (RSSI >= -65 dBm), medium otherwise. Open the events modal with m to browse the last 100; consume them as JSONL via diting monitor.
  • Occupancy quiet vs active — diting calibrate records 5 minutes of "the room is empty" baseline RSSI and writes ./diting-baseline.json. With that file present, the Environment line label becomes quiet / active instead of stable / active. Anything beyond binary occupancy is unreliable on RSSI alone and we deliberately do not implement it.

These are not "Wi-Fi sensing" in the academic sense; they are honest derivative metrics from the data we already have. The Environment line is the live example of what you can responsibly derive from RSSI alone — anything richer (presence behind walls, body pose, vital signs) needs CSI hardware that macOS does not expose.

If you want real Wi-Fi sensing

It's a separate project. Get an ESP32-S3, flash the ESP-CSI toolkit, and visualise the stream. diting's scope ends at "what your Mac's radio sees"; CSI sensing starts where you add external dedicated hardware. We may write a companion repo for that some day, but it will not live inside this codebase.