Description
Determining the number of individuals in a defined area or passing a defined boundary, typically per timestamp or per time window. People counting overlaps heavily with crowd-counting but emphasizes discrete, often gate-style estimation (entries to a room, passengers boarding a bus) rather than continuous density estimation. In wireless settings the inputs are WiFi probe requests, BLE advertisements, CSI variance bursts, or radar Doppler signatures.
Why it's hard
- MAC-address randomization breaks the naive "one MAC = one person" assumption, forcing statistical de-randomization.
- Multi-device users (phone + watch + earbuds) inflate device-based counts; non-carrying users deflate them.
- Counting at a gate requires reliable trajectory direction inference from inherently noisy RSSI.
- Same-room occupancy vs throughput (entries per minute) are different questions answered by the same sensor.
- Privacy regulation (GDPR, ePrivacy) constrains what identifiers may be persisted, even briefly.
Common approaches
- Probe-request capture with statistical de-randomization to reconstruct unique-device counts.
- BLE packet capture with occupancy inference per advertisement bursts.
- CSI-based regression for room-level counts when no device assumption is acceptable.
- Camera-based head detection as the high-accuracy ground-truth baseline.
Source Papers
- ren2023_8cfe ↗ — grouped people counting using mmWave FMCW MIMO radar.
- pronello2025_9dc1 ↗ — low-cost automatic counting at bus stops via WiFi probe requests + deep learning.
- longo2019_b72f ↗ — accurate occupancy estimation with WiFi and BLE packet capture.
- rusca2024_ccca ↗ — privacy-preserving WiFi fingerprint-based people counting.