Description
Estimating whether a space is occupied, by how many people, and increasingly with what activity profile. Unlike instantaneous crowd-counting, occupancy estimation typically targets longer time-scales (minutes to hours) and is framed in the context of building services — HVAC scheduling, lighting control, space utilization analytics. The combined detection / counting / estimation framing is standard in the smart-building literature; this note treats them as one problem with three granularities (binary occupied/empty → integer count → continuous estimate with uncertainty).
Why it's hard
- Sensor sparsity: a single PIR or CO2 sensor per room loses information on multi-occupant dynamics.
- CSI/RSSI counting drift over hours as the environment changes (furniture moves, doors open, multipath shifts).
- Stationary versus moving occupants generate very different signal signatures; stationary occupants are easy to miss.
- Ground truth is hard to label at scale — manual counts, camera baselines, or device-based estimates are each biased differently.
- Cross-room generalization is poor; per-room calibration is the norm.
Common approaches
- WiFi CSI amplitude/phase statistics with classical or deep regressors (per-room training).
- BLE advertisement counts plus occupancy heuristics, sometimes with reinforcement learning.
- Multi-modal fusion of CSI + BLE + CO2 + PIR for building management systems.
- Probe-request / ambient-traffic-based passive monitoring without dedicated active probing.
Source Papers
- chaudhari2024_6efc ↗ — fundamentals, algorithms, and technologies of occupancy detection for smart buildings.
- chen2018_97e0 ↗ — building occupancy estimation and detection: a review.
- jung2025_3f2e ↗ — WiFi CSI-based room-level occupancy estimation in multi-zone residential settings.
- mondal2023_7f7a ↗ — classroom occupancy estimation using WiFi CSI.
- billah2021_69a2 ↗ — BLE Can See: RL approach for RF-based indoor occupancy detection.
- shahbazian2023_1172 ↗ — RF-signal survey of occupancy and activity detection.