Privacy-First Occupancy Analytics with WiFi CSI

Facilities teams need to know how spaces are actually used — peak hours, dead zones, desk utilization — without installing cameras or tracking employee devices. CSI analytics provides anonymous utilization signals from the radio environment.

What to measure

Binary occupancy over time is the robust foundation. Aggregate it:

  • Utilization rate — fraction of time a zone was occupied during business hours.
  • Peak occupancy periods — when does this room actually fill up?
  • Dwell time — how long does occupancy persist once triggered?
  • Zone comparison — which floors, wings, or desks see traffic?

These metrics drive real decisions: rightsizing meeting rooms, adjusting HVAC schedules, reallocating desk inventory. None require knowing who was present.

Temporal aggregation

Raw CSI events are noisy at second-level resolution. Analytics pipelines aggregate over minutes to hours:

  1. Apply debounce/hysteresis at the event layer (same as smart-home automation ).
  2. Bucket into time windows (15-minute, hourly, daily).
  3. Compute utilization = occupied_samples / total_samples per window.
  4. Visualize as heatmaps, trend lines, or zone rankings.

Motion intensity adds a secondary axis — high-traffic vs low-traffic occupancy — when activity detection is available.

Anonymous CSI vs MAC analytics

MAC-address tracking logs device identifiers. Problems:

  • One person, three devices → inflated count.
  • Five people, one shared screen → undercount.
  • Device identifiers may be personal data under GDPR.

CSI analytics reads how bodies disturb the signal. No identifiers are captured or stored. The output is zone-level utilization, not individual tracking. For regulatory context, this is closer to a people-counter beam than a WiFi probe request log.

Honest limits on headcount

Binary occupancy per zone is reliable. Exact person count is not — multipath superposition from multiple bodies creates ambiguous CSI patterns without spatial diversity. WiFree achieved 92.8% crowd-count accuracy with transfer-kernel learning on building-scale CSI, but that required calibrated multi-AP infrastructure, not a single ESP32 node. The detection ladder places person count at rung 7.

Design dashboards around utilization and presence trends, not “47 people on floor 3.” Estimates are possible with multiple nodes and calibration, but treat them as approximate.

Further reading