Description
Building computational representations of how crowds form, move, and dissipate, at granularity ranging from individual agents (microscopic) through density fields (macroscopic) to mesoscopic hybrids. Models are evaluated by how well their simulated trajectories or density evolutions reproduce empirical observations from cameras, GPS traces, BLE/WiFi sensing, or controlled experiments. For this thesis, crowd modeling provides the physics-informed bridge between BLE-derived calibration data and continuous CSI inference: a fluid-style continuum equation acts as a regularization prior.
Why it's hard
- Crowds are heterogeneous (age, group structure, intent) — single-equation models capture only aggregate behavior.
- The same physical layout produces very different crowd patterns under different cultural / situational contexts.
- Calibrating model parameters requires high-quality trajectory or density data that wireless sensors provide only noisily.
- Validation is hard: real disasters cannot be replicated, lab experiments lack realism, simulators encode their own modeling assumptions.
- Coupling between cognition (route choice, panic) and mechanics (collision, push) breaks naive separations.
Common approaches
- Microscopic agent-based models — Social Force, RVO, ORCA — for trajectory-level realism.
- Macroscopic continuum models — Hughes, second-order pedestrian flow PDEs — for density fields.
- Mesoscopic hybrids combining cognitive route choice with continuum dynamics.
- Data-driven surrogates trained on simulated or empirical trajectories.
- Physics-informed neural networks incorporating conservation laws (continuity equation).