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.
1. Scene / layout change
Signal: slow shift in the static multipath fingerprint — mean amplitude and phase profile across subcarriers drift when furniture moves, doors open, or large objects enter the channel.
Fails when: the change is gradual (seasonal furniture rearrangement) and no explicit re-baseline trigger exists.
Wavey: re-baseline trigger — detects environment drift and prompts or auto-resets the empty-room reference.
2. Binary occupancy
Signal: deviation from empty-room baseline — any human body perturbs multipath enough to exceed a change-point threshold. The Berkeley occupancy lineage (FreeDetector, WiFree) uses signal tendency index (STI) — shape similarity between adjacent CSI amplitude curves — rather than a single variance number. Greedy subcarrier selection picks the frequencies most responsive to human presence. WiFree reported 99.1% occupancy detection and 92.8% crowd-count accuracy on building-scale deployments with transfer-kernel learning across rooms. Tolerates noise because the question is binary.
Fails when: baseline is stale, multiple people create ambiguous superposition, or the person is behind heavy shielding.
Wavey: core capability — occupancy events against a learned baseline. See occupancy detection .
3. Motion intensity and trajectory
Signal: broadband energy increase in the 1–5 Hz band; Doppler spread across subcarriers when limbs accelerate. Variance and spectral power rise sharply above baseline.
Fails when: motion is very slow (creeping) or the person is at the edge of coverage with low SNR.
Wavey: motion detection and intensity scoring — the most reliable layer after occupancy. See motion & activity .
4. Activity recognition (HAR)
Signal: distinct time-frequency patterns — but the best systems do not treat HAR as pure pattern matching. CARM (MobiCom 2015) links CSI amplitude dynamics to limb speeds via a quantitative CSI–speed model, then maps body-part speeds to activities. THAT (AAAI 2021) adds dual-stream time-over-channel and channel-over-time features. Both outperform naive LSTM-on-CSI because they respect CSI’s 2D structure.
Fails when: activities are similar (sit vs stand), labels are scarce, or the model was trained in a different room without adaptation — cross-site accuracy can drop from >90% to ~20% without roaming synthesis or adversarial environment stripping.
Wavey: coarse activity classes where labeled data exists; not a general-purpose pose engine. See motion & activity and the activity classifier deep-dive .
5. Micro-presence / respiration
Signal: narrowband periodic modulation at 0.1–0.5 Hz from chest displacement — a micro-Doppler signature far smaller than footstep energy. Requires clean phase, good SNR, and a still environment.
An alternative detection pattern treats the spectrogram as an image: in-car child-presence detection (MobiCom 2024) uses CSI power autocorrelation for macro motion and a spectrogram-enhancement + CNN pipeline for breathing at 5 GHz / 58 subcarriers / 30 Hz — 99% detection, 3% false alarm across 25 vehicle models. The spectrogram-as-image approach generalizes beyond vehicles when SNR and sampling rate are sufficient.
Fails when: range is too great, another person is moving nearby (masking the periodic component), or phase sanitization is insufficient.
Wavey: presence awareness — confirms a still person is in the space. Not a medical vitals device. See presence detection .
6. Fall events
Signal: most published fall detectors treat falls as a classification label — train on fall vs non-fall examples. That assumes falls produce reproducible CSI patterns. They often do not: a fall is an accident, and its impact varies with surface, body orientation, and furniture.
A better framing (SiFall, SenSys 2022): learn the distribution of normal, repeatable activities with an autoencoder (FallNet). Front-end: CSI amplitude → conjugate multiplication denoising → channel dynamics via cosine similarity across subcarrier vectors → acceleration threshold (Θ = 2.5 m/s²) → STFT → FallNet reconstruction. A fall appears as high reconstruction error — an outlier from the person’s routine distribution, not a matched template. Self-supervised incremental learning adapts the normal-activity model online per environment and per user. Reported 98.3% in real-world tests, 94.1% over 3-day continuous monitoring with one false alarm.
Fails when: only one person in a multi-person home (current systems are largely single-room), training lacks real elderly fall samples (most datasets use simulated falls), or routine activities (dropping onto a couch) spike reconstruction error.
Wavey: research direction for elder-care — not a certified medical alert system. See elder-care monitoring .
7. Person count
Signal: superposition of multiple scatterers creates complex interference patterns. With spatial diversity (multiple nodes at different geometry), rough count estimation is possible.
Fails when: people overlap in space, only one TX-RX link exists, or the room geometry creates ambiguous multipath.
Wavey: not a primary target — occupancy yes, exact headcount no. See occupancy analytics for honest scope.
8. Pose / keypoints
Signal: the human body acts as a distributed scatterer — limb positions change which paths dominate. The problem is ill-posed: Person-in-WiFi (ICCV 2019) notes that 9 TX-RX antenna pairs × 30 subcarriers yield ~270 scalar equations per frame to recover a high-dimensional body — so the system learns segmentation masks and joint heatmaps from video labels (OpenPose), not from CSI geometry alone.
Progression in the literature — note the hardware gap at each step:
| System | Hardware | Method | Reported accuracy |
|---|---|---|---|
| Person-in-WiFi (ICCV 2019) | 2× Intel 5300, 3×3 antennas | Multi-task DNN → masks + heatmaps + PAFs | PCK ~78% vs camera 89% |
| WiPose (MobiCom 2020) | Distributed antennas | 3D velocity profile + skeleton prior + RNN | 2.83 cm joint error |
| GoPose (SenSys 2021) | Multi-receiver commodity WiFi | 2D AoA spectrum → CNN spatial + LSTM temporal | 4.5–4.7 cm; works NLOS |
| Person-in-WiFi 3D (CVPR 2024) | Dense WiFi grid, ~4×3.5 m | Transformer set-prediction + Hungarian matching | 91.7 mm (1 person), 125.3 mm (3 people) |
| DensePose-from-WiFi (CMU) | 2 routers, 3×3 antennas | CSI→image modality translation → DensePose-RCNN | UV body regions, multi-person |
The 2.83 cm to 125 mm spread is not contradiction — it reflects different setups, tasks, and person counts. None of these run on a single-antenna ESP32 at 50 pkt/s.
Fails when: commodity hardware (single ESP32 antenna, low packet rate), unconstrained environments, or scenes with more than a few overlapping bodies.
Wavey: not a goal for the open-source ESP32 stack. Pose is demonstrated with research rigs, not $5 nodes.
9. Identity
Signal: gait and habitual motion create person-specific CSI signatures. Two distinct research lines:
- CSI-native identification (C3SL, AAAI 2018) — convex shapelet learning on 3-way CSI tensors as gait fingerprints. 91% accuracy among 20 people on an OpenWrt IoT platform. Requires enrollment and controlled conditions.
- Cross-modal matching (XModal-ID, MobiCom 2019) — simulates WiFi CSI from video 3D mesh, compares gait features to real through-wall CSI magnitude. 85% same-person accuracy without prior CSI training on that person; top-3 ranking 97%. Requires candidate video footage — not identification from CSI alone.
Both are technically feasible in controlled settings; ethically and legally fraught in homes and workplaces.
Fails when: the person changes gait (injury, carrying objects), clothing changes body reflectivity, or regulatory constraints prohibit identification.
Wavey: explicitly not a goal. Wavey detects that someone is present, never who.
10. Through-wall localization
Signal: WiFi penetrates drywall and detects motion in non-line-of-sight conditions — amplitude changes propagate through walls. XModal-ID demonstrated through-wall gait matching at 85% accuracy using magnitude CSI only. Precise position reconstruction behind walls requires dense spatial sampling and calibrated propagation models.
Fails when: walls are concrete or metal-shielded, or the task requires centimeter-level position rather than “motion occurred in this zone.”
Wavey: zone-level motion detection through interior walls — yes. Precise through-wall mapping — no. See security & intrusion .
How to read this ladder
Start at the bottom. Occupancy and motion are engineering problems with known solutions. Presence and HAR need careful preprocessing and often per-site tuning. Everything above person count is research territory or explicitly out of scope for Wavey.
The sensing pipeline explains the processing that makes the lower rungs reliable. Getting started walks through deploying for your first detection task.