Wi-Fi CSI explained: from channel state information to human sensing

By Wu Jun · Updated 2026-11-02 · 6 min read

Wi-Fi CSI (Channel State Information) is the per-subcarrier, per-antenna complex channel measurement taken by a Wi-Fi receiver. Where RSSI gives you one number (how strong the signal is), CSI gives you amplitude and phase for every subcarrier — which is why it can see motion, not just strength.

Everything on this site comes from a real lab: a GL-MT6000 (MT7986) router, 5 GHz / 80 MHz, 256 subcarriers, tx 0/1 × rx 0/1/2/3. A 10-minute run captured 57,293 chain records at 95.5 records/s, which collapse to 7,609 grouped measurements, or about 12.68 Hz; the anomaly-filtered analysis uses 12.64 Hz. The daemon used about 5.2 MB resident memory (raw capture published ^[1]).

A useful mental model

For each measured subcarrier and spatial stream, a simplified receiver model is:

H[k] = Y[k] / X[k]

X[k] is the known training information the receiver can use, Y[k] is what arrived, and H[k] is the estimated channel response. Its magnitude (|H[k]|) is the CSI amplitude; its angle is the CSI phase. A moving person changes the paths through which the radio wave travels, so many H[k] values change together over time. That is the measurable signal behind Wi-Fi sensing ^[2].

This model also explains the main limitation: CSI is a channel measurement, not a camera frame. It does not contain a room image, a face or a direct body map. Research may estimate coarse activity or pose from a controlled multi-antenna capture, but that is an inference from channel changes, not something visible in a raw CSI record.

Four quantities that are not interchangeable

CSI articles often mix records, values, chains and time samples. Keep these units separate:

The 10-minute capture contains both 8-chain and 4-chain groups. Treating all 57,293 records as independent samples, or blindly dividing them by eight, produces the wrong time axis and can create a spectral peak that is not in the physical signal. The public parser groups on chain_info bit 15 instead ^[3].

RSSI vs CSI

What it can and cannot do

Proven directions: presence/occupancy, activity and fall detection, respiration and heart rate (seated, still), gesture recognition, coarse localisation — all surveyed in the field's most-cited review ^[2].

Where projects fail:

1. "Seeing through walls" is over-hyped — you detect channel changes caused by motion, not images.
2. Models do not transfer across rooms, furniture or body poses.
3. Sampling jitter destroys low-frequency analysis (breathing/heart rate).
4. Different chips produce very different CSI — model portability is poor.

Five hardware routes

For long-running, whole-home sensing the router form factor wins: it is already powered 24/7, sits in the middle of the house and needs no extra hardware.

Pipeline

router (capture)                     host (analysis)
OpenWrt + mt76 CSI patch   --UDP-->  parse → buffer → features → detection
mt76-csi-daemon :8888                visualise (heatmap/waterfall/doppler) + storage

Three things are mandatory for CSI to work at all:

1. a CSI-capable firmware (MT7986: the 2026-05-15 build),
2. the frame-type filter (frame_type = 0,34) — without it the firmware reports nothing,
3. the target station's MAC in the MAC filter list.

A 10-minute verification protocol

Before building a classifier, prove that the capture itself is trustworthy:

1. Freeze the radio variables. Record router model, driver and Wi-Fi firmware versions, channel, band width, antenna state, country/regulatory setting and the client model.
2. Create controllable traffic. Use a LAN host and a repeating transfer or ping workload; do not use an internet speed-test server as the packet clock.
3. Capture a known interval. Start with 60 seconds and repeat for 10 minutes. Record the expected start/end time outside the CSI file as well as in it.
4. Check record integrity. Parse every outer record, confirm that timestamps are monotonic enough for the format, and look for long gaps, malformed headers and an unread tail.
5. Group before resampling. Close a measurement on chain_info bit 15. Report the counts of 4-chain, 8-chain and incomplete groups separately.
6. Measure quality, not just volume. Plot group interval, chain completeness, zero-value ratio, RSSI and SNR over time; exclude anomalies only with a documented rule.
7. Publish the chain of evidence. Keep the raw capture, parser version, manifest, SHA-256 and negative results. A polished heatmap without those inputs is not independently reproducible ^[10].

The 10-minute run on this site produced 57,293 outer records, 7,609 complete groups and about 12.68 groups/s. Those three numbers were calculated from the same file but answer different questions; the dataset page exposes the counts and caveats so readers can verify them ^[3].

How to read a CSI claim

Failure map

Start here

• Get 256-subcarrier CSI from a GL-MT6000 router
• Tools, scripts and the 10-minute raw dataset: tools · dataset
• A measured A/B troubleshooting record: why Wi-Fi can slow down after enabling CSI: troubleshooting

Standards and further reading

• IEEE 802.11bf (WLAN Sensing) is the standardisation effort turning sensing into a first-class WLAN capability ^[11].
• A university-run hands-on course (Tsinghua WST) is a good structured starting point for the signal-processing side ^[12].
• Open tools: esp-csi, CSIKit, csiread, SenseFi, ESPectre; community projects such as RuView are lead-only (their headline claims are not peer-reviewed) ^[13].

Please respect local law and ethics: do not use Wi-Fi sensing for non-consensual surveillance. CSI can reveal presence and routines even without images or audio. Obtain consent, minimise retention and publish only anonymised data.