What it is

Device-free crowd-counting WiFi-CSI dataset built explicitly for cross-domain generalisation (privacy-preserving occupancy sensing). Unlike single-actor HAR sets, WiFlow is organised around timed group entry/exit events with event-level ground truth, so it is a direct real-data test bed for hardware-shift-dominates-layout-shift, the cross-environment count mapping (occupancy-csi-variance), and the recalibration story (ble-periodic-calibration).

Verified specs

  • Hardware: two ESP32-WROOM-32U boards (single external antenna each) as Tx/Rx, a Raspberry Pi collector. This is a third commodity-CSI platform distinct from Intel-5300 (wimans) and Nexmon-80 MHz (meneghello-80mhz-csi) — valuable for isolating the hardware-shift axis.
  • CSI: sampled at 100 Hz, 52 effective subcarriers / packet.
  • Environments: Laboratory (7 m × 7 m, concrete/drywall) and Classroom (5.5 m × 9 m, linoleum).
  • Occupancy: 10 volunteers in controlled groups of 2 / 5 / 9, with timed entry/exit.
  • Scale: ~6 h raw CSI → ~24,000 labelled windows (after event-purity filtering).
  • Cross-domain: 6 transfer conditions = {location} × {group size} (e.g. Lab-2 → Classroom-5).

Acquisition & storage

  • Access: proprietary but provided to academic researchers on request; governed by the IEEE LPPL v1.3. Not yet acquired — queued alongside the Tier-B public-data pull.
  • Landing: source_url above (topic overview). Request access via the authors.

Why it matters here

The ESP32 platform + genuine 2/5/9-person groups + labelled cross-domain splits make WiFlow the cleanest public complement to the two datasets we already hold: it adds a third hardware class to the A2 hardware-vs-layout contrast (csi_hardware_vs_layout.py) and a proper cross-environment count benchmark that is not confounded by band/subcarrier differences the way the WiMANS↔Meneghello contrast is.

Namesake disambiguation

A different project also uses the name WiFlow — a WiFi pose-estimation network (spatio-temporal feature decoupling, repo https://github.com/DY2434/WiFlow-WiFi-Pose-Estimation-with-Spatio-Temporal-Decoupling), whose single-actor CSI→2D-pose dataset is documented separately in wiflow-pose. It is unrelated to this crowd-counting dataset; do not conflate them when citing.