Description
Estimating the number of people present in a monitored area from sensor observations, without requiring each individual to carry a cooperating device. In wireless-sensing literature the canonical setting is device-free counting: bodies perturb the propagation environment of an existing WiFi or BLE link, and the count is regressed from CSI amplitude/phase statistics or RSSI variance. Crowd counting is the entry-point problem of indoor crowd modeling — every higher-order task (density, flow, dynamics) reduces to a counting question on a sub-region.
Why it's hard
- Multipath in indoor environments makes the mapping from "people present" to "signal disturbance" highly non-linear and site-specific.
- Ground truth is expensive: cameras violate privacy, manual counts don't scale, BLE-tagging requires cooperation.
- The same signal change can come from one person walking close to the link or three people far away; counting and localization couple.
- Saturation effects — once the channel is fully scrambled, additional bodies stop adding distinguishable variance.
- Models trained in one room rarely transfer to another (see environment-dependence).
Common approaches
- CSI amplitude/phase statistics fed to classical regressors or CNNs/LSTMs.
- WiFi probe-request / BLE advertisement counting (device-based, undercounts due to MAC randomization).
- mmWave FMCW radar with point-cloud clustering for grouped counts.
- Camera-based density regression as a comparison upper bound.
- Few-shot / domain-adaptive learning to reduce per-site labeling burden.
Source Papers
- hou2023_bf83 ↗ — DASECount: domain-agnostic few-shot wireless indoor counting.
- choi2022_17c2 ↗ — Wi-CaL: WiFi sensing + ML for device-free crowd counting and localization.
- sakhnini2024_de9b ↗ — mmWave FMCW radar counting in a festival setting.
- golammowla2024_1b1f ↗ — CSI-based people counting via occupancy detection cues.
- rusca2024_ccca ↗ — privacy-preserving WiFi-fingerprint counting for crowd management.