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.

Activity recognition (HAR) has three generations in the literature:

Generation 1: physics-linked models (CARM, MobiCom 2015)

Before deep learning dominated, CARM built quantitative models:

CSI–speed — amplitude change frequency maps to limb speed via Doppler (path-length change of one wavelength = 2π phase shift).
CSI–activity — activities differ by which body parts move at which speeds (legs periodic in walking, torso impulse in falling).

Features: PCA denoising, discrete wavelet transform, HMM per activity for state transitions. >96% on COTS WiFi using amplitude, not phase — CFO on commodity hardware makes raw phase unusable for macro-motion HAR. Still a strong baseline because features encode mechanism, not just correlation.

Generation 2: spectrogram CNNs

STFT tiles (time × frequency) fed to CNNs. Learns motion textures without explicit speed models. Works when you have labels; brittle cross-room without adaptation.

Generation 3: dual-stream transformers (THAT, AAAI 2021)

CSI has two informative axes:

Stream	What it captures	Missed by LSTM-only
Channel-over-time	How each subcarrier evolves	—
Time-over-channel	Subcarrier correlation at one instant	✓

THAT’s Multi-scale Convolution Augmented Transformer (MCAT) processes both streams with Gaussian range encoding. +2.2 pts accuracy over ABLSTM, 1.8–3.4× faster — because a single temporal sample is too elementary to represent a CSI pattern.

The cross-room transfer problem

CrossSense (MobiCom 2018) showed gesture recognition accuracy collapse from >90% in-room to ~20% cross-site with a single global model. Their fix: a roaming model that synthesizes target-environment training data from a small calibration set, plus a mixture-of-experts runtime selector (DTW-based) that routes signals to specialized models.

Separately, the EI framework (MobiCom 2018) uses adversarial training — feature extractor vs domain discriminator — to strip environment-specific information before classification.

Without one of these (or per-site fine-tuning), do not deploy a model trained elsewhere. See the sensing pipeline for the full adaptation landscape.

What Wavey targets

Wavey emits motion events (intensity, duration) as the primary output. Coarse activity classes are supported where labeled data from the deployment environment justifies a classifier. Multi-person 3D pose (CVPR 2024: ~92–125 mm joint error with dense WiFi nodes in a 4×3.5 m rig) is research territory — not ESP32 territory. See the detection ladder , rung 4.

For building a classifier from scratch, read building an activity classifier on CSI .

WiFi Motion & Activity Detection