Wavey | WiFi CSI Sensing System on Wavey

Device-Free Occupancy Detection with WiFi CSI

Sat, 20 Jun 2026 12:00:00 +0000

The most useful sensing question is also the easiest to answer reliably: is anyone here? Binary occupancy asks for a single bit — empty or occupied — which tolerates noise far better than counting people or classifying pose.

Change-point detection on a baseline

Wavey learns an empty-room baseline: the mean amplitude profile, typical subcarrier variance, and phase structure when no one is in the space. Live CSI compared against this reference uses change-point detection — when the deviation exceeds a threshold sustained over a window, occupancy is declared.

WiFi Sensing vs Cameras for Occupancy & Motion

Sat, 20 Jun 2026 12:00:00 +0000

Cameras and WiFi CSI solve overlapping but not identical problems. The choice is about what data you need from the sensor, not just privacy preferences.

Sensing mechanism

	Camera	WiFi CSI
Signal	Visible light (photons)	Radio channel (amplitude + phase per subcarrier)
Output	Image frames (pixels)	Channel measurements (complex vectors)
Identity	Trivial (face recognition)	Not available by design
Pose	Direct (keypoint models)	Research frontier; not on commodity ESP32

A camera answers “what does this look like?” CSI answers “how did the radio channel change?”

WiFi Motion & Activity Detection

Sat, 20 Jun 2026 12:00:00 +0000

Movement is the strongest signal in CSI. Limb acceleration creates broadband energy in the 1–5 Hz band, variance spikes across subcarriers, and the channel changes within a single packet window. Detecting that something moved is straightforward. Classifying what happened is where a decade of published systems disagree on approach.

From motion to activity

Motion detection is a threshold on variance or spectral power — binary, fast, environment-agnostic within a single deployment.

WiFi CSI Sensing vs mmWave Radar

Sat, 20 Jun 2026 12:00:00 +0000

Both are non-visual RF sensors. The tradeoff is dedicated ranging hardware vs commodity WiFi reuse.

RF front-end

mmWave radar uses a purpose-built transceiver (typically 60 GHz) with chirp modulation designed for range-Doppler processing. Resolution: centimeter-range distance, sub-Hz Doppler for micro-motion.

WiFi CSI uses the existing 2.4/5 GHz WiFi PHY. Resolution: limited by wavelength (~6 cm at 5 GHz) and packet-rate sampling. Fine-ranging is not the design goal.

Fine-motion SNR

Radar front-ends are optimized for detecting sub-millimeter chest displacement — breathing and heartbeat are first-class signals. ESP32 CSI can detect breathing after phase sanitization, but the hardware noise floor is higher and SNR is placement-dependent.

Presence & Breathing Detection with WiFi CSI

Sat, 20 Jun 2026 12:00:00 +0000

The hardest sensing case for conventional sensors is a person who is not moving. PIR reports empty. Cameras work but raise privacy concerns. CSI can detect micro-presence — a still body whose chest rises and falls with respiration.

Micro-Doppler physics

Breathing displaces the chest wall by a few millimeters at 0.1–0.5 Hz (6–30 breaths per minute). This creates a micro-Doppler modulation on the CSI phase — a narrowband periodic signal far weaker than footstep energy but structurally distinct from an empty room.

WiFi CSI Sensing vs PIR Motion Sensors

Sat, 20 Jun 2026 12:00:00 +0000

PIR and WiFi CSI both detect people without cameras. They measure completely different physical quantities.

Sensing mechanism

PIR detects changes in infrared radiation within a ~120° cone. A motionless person emits constant IR — no change, no trigger. After a timeout, the sensor reports empty. This is the “lights off while I’m sitting still” failure mode — a fundamental property of differential IR sensing, not a firmware bug.

WiFi CSI detects changes in radio channel state across the whole room. A still person modulates phase via breathing (micro-Doppler at 0.1–0.5 Hz). Macro-motion variance may be near zero while periodic phase modulation confirms presence.

Elder-Care Monitoring with WiFi Sensing

Sat, 20 Jun 2026 12:00:00 +0000

Aging in place needs continuous awareness without cameras in bedrooms and without wearables that get left on the nightstand. WiFi CSI offers ambient monitoring — presence, movement patterns, and fall-like events — from nodes already in the home.

Why fall is not a classification problem

Most early RF fall detectors train a classifier: fall vs walk vs sit. That assumes falls produce reproducible CSI signatures you can label and match. In practice, falls are accidents — impact varies with surface, body angle, and what they hit. Labeling enough real elderly falls to train a classifier is ethically and practically hard; most training data is simulated.

WiFi CSI Sensing vs Wearables & Phone Tracking

Sat, 20 Jun 2026 12:00:00 +0000

Wearables and phone tracking sense devices. WiFi CSI senses bodies. The coverage and data models are fundamentally different.

What gets detected

	Wearable / phone tracking	WiFi CSI
Signal source	BLE beacon, UWB tag, WiFi probe/MAC	Body perturbation of radio channel
Requires cooperation	Yes — person carries device	No — device-free
Misses	Anyone without the device	No one in RF coverage
Identity	Device ID (often personal data)	Anonymous — no identifiers

A guest, child, intruder, or elder who forgot their pendant is invisible to device tracking. CSI detects any body that perturbs the channel.

WiFi Sensing for Security & Intrusion Detection

Sat, 20 Jun 2026 12:00:00 +0000

Security sensing needs to work when lights are off, cameras are blind, and the intruder carries nothing. WiFi CSI detects bodies by how they disturb the radio field — device-free, darkness-invariant, and effective through many interior walls.

NLOS motion through building materials

WiFi at 2.4 and 5 GHz penetrates drywall, plaster, and wooden doors. An intruder moving in an adjacent room perturbs multipath paths that traverse the wall — amplitude shifts propagate to the receiver even without line of sight.

Smart-Home Automation with WiFi Presence Sensing

Sat, 20 Jun 2026 12:00:00 +0000

The most common smart-home failure is lights switching off while you are sitting still. The root cause is using motion pulses where presence holds are needed. WiFi CSI provides the hold signal.

Presence-hold vs motion-pulse

Event type	Semantics	Good for
Motion pulse	Something moved, now	Triggering on entry, security alerts
Presence hold	Someone is here, still	Keeping lights on, HVAC running
Absence clear	Space is empty	Shutting down scenes, energy savings

PIR emits motion pulses. When motion stops, the pulse ends and the automation assumes empty. CSI occupancy and presence detection emit holds — the space remains “occupied” while a still person breathes in it.

Privacy-First Occupancy Analytics with WiFi CSI

Sat, 20 Jun 2026 12:00:00 +0000

Facilities teams need to know how spaces are actually used — peak hours, dead zones, desk utilization — without installing cameras or tracking employee devices. CSI analytics provides anonymous utilization signals from the radio environment.

What to measure

Binary occupancy over time is the robust foundation. Aggregate it:

Utilization rate — fraction of time a zone was occupied during business hours.
Peak occupancy periods — when does this room actually fill up?
Dwell time — how long does occupancy persist once triggered?
Zone comparison — which floors, wings, or desks see traffic?

These metrics drive real decisions: rightsizing meeting rooms, adjusting HVAC schedules, reallocating desk inventory. None require knowing who was present.

IEEE 802.11bf and the WiFi Sensing Privacy Gap

Fri, 31 Jul 2026 12:00:00 +0000

IEEE 802.11bf was ratified in September 2025. It standardizes WiFi sensing — making CSI-based occupancy, motion, and activity detection a first-class capability in commercial WiFi hardware. For the sensing industry, this is a milestone. For RF privacy, it is an escalation: the attack surface just became standardized.

This post covers what 802.11bf enables, what privacy protections it includes (and omits), and what organizations evaluating WiFi sensing should demand from vendors.

RF Privacy Defenses: Obfuscation, Metasurfaces, and Their Limits

Thu, 30 Jul 2026 14:00:00 +0000

The defense literature for unauthorized WiFi sensing is almost entirely about preventing sensing, not detecting surveillors. This post surveys the major approaches, what each achieves in published evaluations, and where the practical limits sit for real deployments.

For the attacks these defenses target, see WiFi sensing attacks . For detection alternatives, see detecting unauthorized RF sensing .

Two defense philosophies

Defenses fall into two categories that mirror the attack taxonomy:

Type	Mechanism	Who acts
Active defense	Manipulate transmitted signals so eavesdroppers get wrong CSI	Transmitter, AP, or dedicated obfuscation device
Passive defense	Third party alerts users or protects sensing results	Monitoring system, RF privacy platform

Active defense is mature in research — multiple papers with real-world evaluations. Passive defense (detection and alerting) is barely explored. An RF privacy platform needs both.

Detecting Unauthorized RF Sensing: A Layered Approach

Thu, 30 Jul 2026 12:00:00 +0000

If someone is using WiFi signals to observe your space, can you detect them? The answer depends heavily on what kind of attacker — active emitters are tractable; pure passive receivers may be undetectable. This post maps five detection layers, assesses each against published research, and identifies what is shippable on commodity hardware today.

For the attacks these layers defend against, see WiFi sensing attacks . For the obfuscation-based alternative, see defenses and limits .

WiFi Sensing Attacks: LeakyBeam, BFId, and the BFI Side Channel

Wed, 29 Jul 2026 12:00:00 +0000

Two papers published in late 2025 moved WiFi sensing from a research curiosity to a documented privacy threat at top security venues. Both exploit the same side channel: Beamforming Feedback Information (BFI) — plaintext packets that WiFi 5 and 6 clients transmit to help access points steer signals.

This post summarizes the attacks, what they demonstrate, and why they matter for RF privacy research.

What BFI is and why it leaks

WiFi 5 (802.11ac) introduced explicit beamforming: the AP sends sounding frames, the client measures the channel, and the client responds with compressed beamforming feedback describing how the signal arrived at each antenna. The AP uses this to compute a steering matrix for directional transmission.

RF Privacy: The Flip Side of WiFi Sensing

Tue, 28 Jul 2026 12:00:00 +0000

Wavey reads WiFi Channel State Information to infer occupancy, motion, and presence — device-free and camera-free. That is the legitimate use case. The same physics works in reverse: anyone who can observe the radio environment can infer the same things about you, often without your knowledge and without specialized hardware.

This post is the first in a research series on RF privacy — detecting, auditing, and mitigating unauthorized RF sensing. It covers why the problem exists, what current wireless security products miss, and where the research frontier sits.

Privacy Policy

Thu, 23 Jul 2026 00:00:00 +0000

Effective date: 23 July 2026

This Privacy Policy explains how the Wavey project team ("Wavey", “we”, “us”, or “our”) collects, uses, discloses, and otherwise processes information when you visit waveyhq.dev (the “Site”) or submit feedback through it.

The Wavey project team operates the Site from India. For privacy questions or requests, contact mail@waveyhq.dev .

1. Scope

This Policy applies only to the public Site. It does not govern:

the Wavey Console at console.waveyhq.dev ;
data collected by a Wavey software or hardware deployment;
third-party websites, repositories, communities, or services linked from the Site; or
information processed by a person or organization that deploys Wavey independently.

Those products and services may have separate terms and privacy notices. A person or organization deploying WiFi sensing is independently responsible for its data practices and legal obligations.

Terms of Use

Thu, 23 Jul 2026 00:00:00 +0000

Effective date: 23 July 2026

These Terms of Use ("Terms") govern your access to and use of waveyhq.dev (the “Site”), operated by the Wavey project team ("Wavey", “we”, “us”, or “our”).

Please read these Terms carefully. By accessing or using the Site, you agree to these Terms to the extent permitted by applicable law. If you do not agree, do not use the Site.

1. Scope

The Site publishes documentation, technical articles, project information, public discovery metadata, and links relating to Wavey, an open-source WiFi Channel State Information (“CSI”) sensing project.

WiFi CSI Detection Ladder: Tasks, Feasibility, and Limits

Fri, 26 Jun 2026 12:00:00 +0000

Not every sensing task is equally reachable with commodity WiFi CSI. This page is the authoritative map: what each detection level requires in the signal, what breaks it in practice, and where Wavey sits on the ladder.

For how signals are processed before inference, see the sensing pipeline . For the underlying physics, see how it works .

The ladder

Detection difficulty increases as you move from coarse environment change to fine per-person detail. Each level builds on the ones below it.

WiFi CSI Sensing Pipeline: Signal Processing and ML

Fri, 26 Jun 2026 12:00:00 +0000

Raw CSI from a commodity WiFi link is not inference-ready. The gap between a research demo and a deployment that works in someone’s living room is almost entirely signal processing and honest ML engineering. This page covers that layer — what happens between packet capture and a detection event.

For the physics of why bodies show up in CSI, see how it works . For what tasks are feasible at each difficulty level, see the detection ladder .

Commodity CSI: Capture Limits and Calibration

Wed, 24 Jun 2026 12:00:00 +0000

Research papers often use USRPs, Intel 5300 NICs (3×3 MIMO, 9 antenna pairs), or dense WiFi grids. Person-in-WiFi used 2× Intel 5300 sets with 30 subcarriers per link — 270 scalar measurements per frame. WiVi pushed 100 pkt/s with custom OpenWrt firmware on TP-Link routers, producing 100×114 CSI frames. Wavey uses a $5 ESP32. This post is about what that hardware actually delivers, where it falls short, and when it is enough.

Baselines, Drift, and Generalization

Mon, 22 Jun 2026 12:00:00 +0000

Every occupancy system depends on knowing what “empty” looks like. This post covers how baselines are built, how they drift, why models trained in one room fail in another, and when to reset.

For the detection mechanism, see occupancy detection . This is the stability engineering layer.

Building an empty-room fingerprint

During a known-empty period (nights, scheduled away times, or initial setup), collect CSI for 5–30 minutes. Compute:

Mean amplitude profile per subcarrier: $\bar{A}[k] = \text{mean}_t(|CSI[k,t]|)$
Variance envelope per subcarrier: typical fluctuation range when empty
Phase reference (sanitized): mean phase structure after CFO/SFO removal
Shape similarity (STI) — for each subcarrier, how much does the amplitude curve’s shape change when someone enters? Greedy subcarrier selection picks the most responsive frequencies. This is the FreeDetector / WiFree approach; WiFree reported 99.1% occupancy on building-scale deployments.

This fingerprint is the baseline. Live CSI compared against it powers change-point detection: if the current window’s statistics deviate beyond the empty-room envelope, declare occupied.

WiFi CSI Sensing FAQ

Sat, 20 Jun 2026 12:00:00 +0000

Practical questions about WiFi CSI sensing and Wavey. For technical depth, see how it works , the sensing pipeline , the detection ladder , and technical deep-dives .

Getting Started with Wavey

Sat, 20 Jun 2026 12:00:00 +0000

Wavey is an open-source WiFi CSI sensing system. This page covers deployment architecture — for signal processing detail, see the sensing pipeline . For what to detect first, see the detection ladder .

Setup steps live in the GitHub repository — they change as the project matures.

Three-layer architecture

[ ESP32 CSI nodes ] -> [ Python pipeline ] -> [ console / automations ]
 capture CSI preprocess + features visualize + emit events

Sensing nodes (ESP32). Capture CSI from ambient WiFi traffic. See commodity CSI limits for hardware constraints.
Analysis pipeline (Python). Outlier rejection, amplitude features (primary on ESP32), phase sanitization for micro-motion, baseline learning, inference. See the sensing pipeline for the processing stack.
Console and integrations. Live visualization at console.waveyhq.dev ; events for automation and analytics .

Before you begin

One or more CSI-capable ESP32 modules (any ESP32 family chip)
A WiFi environment with traffic the nodes can sniff
A host machine for the Python stack

Pick your first detection task

Start at the bottom of the detection ladder — the most robust tasks first:

WiFi Sensing Glossary

Sat, 20 Jun 2026 12:00:00 +0000

Definitions for terms used across the Wavey documentation. For how they connect, start with how it works , the sensing pipeline , and the detection ladder .

Channel State Information (CSI)

A per-subcarrier measurement of how a WiFi signal traveled from transmitter to receiver — amplitude and phase at each OFDM frequency slice. The core sensing signal. See how it works .

RSSI (Received Signal Strength Indicator)

A single scalar: total received power. Too coarse for micro-motion or activity recognition. CSI preserves the per-subcarrier detail RSSI discards.

How WiFi CSI Sensing Works

Sat, 20 Jun 2026 12:00:00 +0000

Every WiFi link already measures the radio channel between transmitter and receiver. Channel State Information (CSI) is that measurement, kept at full resolution: amplitude and phase across dozens of frequency slices. When a body moves in the space, it perturbs those measurements in predictable ways. This page explains the physics. For signal processing and ML, see the sensing pipeline . For what can be detected, see the detection ladder .

CSI vs RSSI

RSSI is one number — total received power. It answers “how strong is the signal?” and nothing else.

Building an Activity Classifier on CSI

Thu, 18 Jun 2026 12:00:00 +0000

Detecting that something moved is a threshold problem. Classifying what happened is a machine learning problem. This post walks through building an activity classifier on CSI — from the simplest baseline to a spectrogram CNN — and explains why the hard part is not the model.

For the use-case framing, see motion & activity detection . This is the engineering walkthrough.

Level 0: variance threshold

Compute subcarrier variance over a 1-second sliding window. If variance exceeds a threshold tuned on empty-room data, declare motion. No labels needed. No ML framework needed.

About Wavey

Sat, 13 Jun 2026 12:00:00 +0000

Wavey is an open-source RF sensing system that uses WiFi Channel State Information (CSI) to observe physical environments - occupancy, motion, and presence - without cameras, wearables, or active participation from the people being sensed.

Our mission

Spatial awareness usually means installing cameras (a privacy problem), buying expensive radar or LiDAR (a cost problem), or asking people to carry phones and tags (a friction problem). Wavey takes a different path: it reads the WiFi signals that already fill our spaces and turns their tiny disturbances into useful information.

Contact Us

Sat, 13 Jun 2026 12:00:00 +0000

If you have questions, contributions, or ideas for Wavey, we’d love to hear from you!

Discord: Hang out with us!
GitHub: Wavey Project
GitLab: WaveyHQ
Email: mail@waveyhq.dev

The WiFi Channel as a Distributed Sensor

Sat, 13 Jun 2026 12:00:00 +0000

Most explanations of WiFi sensing stop at “CSI measures amplitude and phase.” This post goes one level deeper: what a single CSI snapshot actually contains, how multipath superposition works as a concrete model, and why collapsing that into RSSI throws away the information you need.

For the full physics treatment, see how it works . This is the worked-example version.

One packet, one matrix

When an ESP32 receives a WiFi frame, the PHY layer reports CSI as a complex vector: