Developing methods for real-time pedestrian agent-based modelling: An ensemble Kalman Filter approach

This research aims to progress the development of real-time pedestrian simulation methods. Simulating pedestrian systems in close to real-time affords us up-to-date knowledge of how a population is distributed around a city. This may allow local authorities to react to demand for services that varies in time and space by engaging in data-driven decision-making approaches.

One of the most popular approaches for simulating pedestrian motion is Agent-Based Modelling. Such models are typically calibrated using historical data to ensure that they reflect the real system. Ultimately, however, these models usually incorporate some degree of randomness to emulate the variability of human behaviour resulting in the model diverging from the real system. This restricts their application to predominantly offline simulation.

Another potential approach to exploring pedestrian systems is the use of Big Data. Data are being generated in increasing volumes at an increasing velocity, and data regarding pedestrian flows are no exception. Such data, however, may offer sparse coverage with regards to time, space or demographic portions of the population — an issue from which Agent-Based Modelling does not suffer. Consequently, we seek an approach that allows us to benefit from both simulation and data. At present, there is not a consensus on how best to incorporate new observations into an Agent-Based Model whilst it is running.

The challenge of incorporating observations into running models is frequently tackled in other fields such as numerical weather prediction where meteorologists make use of data assimilation techniques to introduce observations into their models. This thesis, therefore, focuses on adapting the Ensemble Kalman Filter data assimilation technique to work with Agent-Based Models of pedestrian motion. Given the differences between Agent-Based Models and the models to which data assimilation methods are typically applied, additional challenges arise. This thesis will document some of these challenges and propose solutions.