Description
Data assimilation combines a prior model (forecast from a physics or agent-based simulator) with sparse observations to produce a corrected posterior state in real time. Variational (3D-Var, 4D-Var) and Bayesian (particle-filter, ensemble Kalman) flavours dominate. In the thesis context, data assimilation is the canonical machinery for fusing BLE-anchored ground-truth trajectories with continuum / agent-based crowd models — turning periodic calibration campaigns into a continuously corrected state estimate.
When it's used
- BLE-CSI fusion under a physics prior
- Real-time correction of
agent-based-modelruns from sparse observations - Drift detection in deployed sensing systems
- Hindcasting crowd states from lagged observations
Limitations
- Computationally heavy at scale
- Requires both a credible prior model and well-characterised observation noise
- Posterior collapse if observation precision is over-trusted