Description
Estimating a continuous-or-grid density field over a monitored area, rather than a single integer count. Density estimation is the spatial generalization of crowd-counting and the input most modeling pipelines actually need: pedestrian-flow models consume density and velocity fields; crowd-safety thresholds are defined per-zone. In the BLE-calibrated CSI thesis, density-field reconstruction from CSI between calibration campaigns is the headline inference task.
Why it's hard
- Wireless sensors provide highly spatially-aggregated readings; recovering a fine grid requires spatial priors.
- Density discontinuities (queue boundary, room edge) are hard to recover from smooth-by-default models.
- Saturation: high-density regions all look the same to a single CSI link.
- Validation requires per-cell ground truth at high spatial resolution, which is costly.
- Mass-conservation between time steps is a useful prior that off-the-shelf regressors break.
Common approaches
- CNN density-regression models (vision baseline) with pixelwise Gaussian targets.
- Grid-based interpolation from a sensor-cell occupancy graph.
- Kernel density estimation from device-trajectory snapshots (BLE/WiFi probe).
- Physics-informed density inference enforcing continuity equation.
Source Papers
- sindagi2018_e579 ↗ — CNN-based single image crowd counting and density estimation.
- alam2022_0e15 ↗ — estimating indoor crowd density and movement using WiFi sensing.
- di2023_285b ↗ — physics-informed deep learning for traffic-state estimation (related framing).
- bendalibraham2021_476e ↗ — recent trends in crowd analysis.