Fluid dynamics-Like Open-source Crowd Knowledge-driven Simulator (FLOCKS)

As urban areas become more densely populated, effectively managing and predicting pedestrian dynamics is critical in areas such as public safety, architecture and transportation. For example, crowd managers could quickly assess different escape routes during an evacuation to minimise injuries; architectural design can be optimised by considering the dynamics of pedestrian movement; and the development of autonomous vehicles could benefit from a better understanding of how people navigate crowded urban spaces.
However, current simulation tools lack the ability to model large, dense crowds in real time. On the one hand, the agent-based approach, in which pedestrians are modelled individually, is so computationally expensive that it makes real-time crowd management impossible. On the other hand, current models that liken crowds to continuum 'thinking fluids' are severely limited in their accuracy because, in order to obtain a closed system of equations, they typically make simplistic assumptions about constitutive relationships, such as that relating the local speed of pedestrians at a point to the state of the crowd in the area around that point.
This proposal will derive the first fluid dynamics-like crowd model using a data-driven modelling technique to combine the scalability of the continuum description with the accuracy of data-driven approaches. As the opaque 'black box' models provided by most machine learning methods are not suitable for high stakes applications such as those involving crowds, a novel machine learning approach will be developed that extends the current ability to distil mathematical expressions from the data. We will use two complementary data sources - firstly, surrogate pedestrian models, as these provide clean datasets from controlled experiments that are ideal for developing the machine learning tools, and secondly, real crowd observation data to ensure the accuracy of the distilled model. Integrating this model into OpenFOAM (a widely...