Skip to main content
Movement is the strongest signal in WiFi CSI. Limb acceleration creates broadband energy in the 1–5 Hz band, variance spikes across subcarriers, and channel changes within a single packet window. Wavey detects that something moved with a simple threshold, and supports coarser activity classification where labeled data from the deployment environment justifies a classifier.

​
From motion to activity

Motion detection is a threshold on variance or spectral power — binary, fast, and environment-agnostic within a single deployment. Human activity recognition (HAR) has three generations in the published literature. Each generation trades model complexity against cross-environment robustness.
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. Greater than 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.
STFT tiles (time × frequency) fed to CNNs. Learns motion textures without explicit speed models. Works when you have labels; brittle cross-room without adaptation.
CSI has two informative axes: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 greater than 90% in-room to approximately 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: approximately 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.

​
Further Reading

Detection Ladder

HAR feasibility and limits

Security & Intrusion

Motion as a zone alarm

WiFi Motion Detection Without Camera

Building an activity classifier from scratch

Getting Started

Set up your first Wavey deployment