Case study · Network & telemetry AI
On-device WiFi sensing
Human-activity recognition from WiFi channel state information (CSI) — no cameras — running fully offline on an ARM Cortex-A with 1 GB of RAM and no GPU. A compact CNN + BiLSTM + attention network classifies seven activities with 93.5% test accuracy.
- 93.5%test accuracy67,439 samples across 7 activity classes
- ~230 KBmodel sizeFits comfortably on embedded Linux devices
- No GPUrequiredARM Cortex-A · 1 GB RAM · offline inference
The challenge
WiFi signals change as people move, so a router can sense activity without a camera — a privacy advantage in homes and care settings. But the model has to live on the device: no GPU, 1 GB of RAM, no cloud round-trip.
Our approach
A small hybrid network: a CNN layer extracts local patterns from the CSI stream, a BiLSTM captures how they evolve over time, and attention weights the most informative moments before a fully connected classifier. The result is a ~230 KB model that runs offline on embedded Linux.
Workflow
How it works, step by step.
The pipeline in 5 stages, from capture to classify. Each stage has a clear input, a clear output and a reason to exist.
- 01
Capture
Channel state information streamed from the WiFi chipset.
- 02
CNN
32 filters extract local spatial features from each CSI window.
- 03
BiLSTM
64 hidden units capture temporal dependencies in both directions.
- 04
Attention
Weights the most informative time steps.
- 05
Classify
A fully connected layer outputs one of seven activities, on-device.
Results
What it produced.
Per-class F1 · test set
Target hardware
- Processor
- ARM Cortex-A
- Memory
- 1 GB RAM
- Accelerator
- None — CPU only
- Model size
- ~230 KB
- Runtime
- Embedded Linux · offline inference
Static postures are near-perfect; transitions (get up, get down) are the hardest classes — which is where further data collection pays off.
Stack & techniques
- CNN
- BiLSTM
- Attention
- WiFi CSI
- Embedded Linux
- ARM Cortex-A
Delivered by members of the Saerosense founding team. Client names are withheld unless public — references are available on request.
Related service
Start a project
Have a similar problem?
Send a few lines about your data and the outcome you want. We reply within two business days with a feasibility read and a proposed scoping plan.