Bayesian Behavioural Model Estimation for Live Crowd Simulation

The increasing availability of real-time crowd observation data enables the development of agent-based crowd simulations that incorporate real-time data feeds, i.e. live crowd simulations, which achieve accurate crowd forecasting for real-time interventions for improved and safer mobility. Various approaches for live crowd simulations have recently been proposed. However, existing methods cannot offer long forecasting lead times, which are crucial for planning and implementing timely interventions. To address this issue, we develop a Bayesian behavioural model estimation for live crowd simulations that sequentially estimates the underlying behavioural model assumed behind the observed crowd flows. In real crowds, although apparent behaviours change over time, an invariable rule that determines behaviour often exists behind the crowd. The developed method estimates the underlying invariable behavioural model and provides reliable long-term crowd flow forecasting. The experimental results show that the developed method can accurately forecast long-term crowd flows using aggregate observations, whereas the state-of-the-art forecasting method fails to provide reliable forecasting when the apparent behavioural tendency changes. We also demonstrate that the developed method can provide forecasting results that are sufficient to consider interventions, even in the presence of a data-model mismatch.