Description
The computational realization of a crowd-modeling specification — turning equations or rules into time-stepped trajectories that can be visualized, evaluated, or coupled to building-design tools. The simulation problem is distinct from the modeling problem because it adds numerical, parameter-estimation, and validation concerns: which simulator, which integrator, which calibration data, which fitness function. For the thesis, crowd simulation is what consumes BLE-calibrated parameters and exposes them as predictions for CSI-only inference between calibration campaigns.
Why it's hard
- Parameter spaces are large and weakly identifiable from observation.
- Real-time simulation under closed-loop sensor feedback is constrained by integration cost.
- Validation metrics (Hausdorff trajectory distance, density-RMSE, fundamental-diagram fit) all measure different things and disagree.
- Cross-simulator portability is poor — Vadere, MomenTUMv2, JuPedSim, AnyLogic each encode different microscopic assumptions.
- Calibration drift over time is rarely accounted for in published simulators.
Common approaches
- Particle filters fitting agent-based models to trajectory data.
- Bayesian / surrogate-based parameter estimation for expensive simulators.
- Differentiable physics for end-to-end-learnable agent kinematics.
- Open-source simulator ecosystems (Vadere, JuPedSim, PEDSim) with shared scenario libraries.
Source Papers
- wolinski2014_f409 ↗ — parameter estimation and comparative evaluation of crowd simulations.
- malleson2020_7b38 ↗ — real-time agent-based simulation with particle filter.
- yang2020_e295 ↗ — review on crowd simulation and modeling.
- duives2013_3924 ↗ — state-of-the-art crowd motion simulation models.
- kleinmeier2019_e6cd ↗ — Vadere open-source simulator.