Understanding pedestrian dynamics using machine learning with real-time urban sensors

Quantifying, understanding and predicting the number of pedestrians that are likely be present in a particular place and time (‘footfall’) is critical for many academic, business and policy questions. However, limited data availability and complexities in the behaviour of the underlying pedestrian ‘system’ make it extremely difficult to accurately model footfall. This paper presents a machine learning model that is trained on a combination of hourly footfall count data from sensors across a city as well as important contextual factors that are associated with pedestrian movements such as the structure of the built environment and local weather conditions. The aims are to better understand the relationship between various contextual factors and footfall and to predict footfall volumes across a spatially heterogeneous city. The case study area is the city of Melbourne, Australia, where abundant pedestrian count data exist. Time-related variables, particularly time-of-day and day-of-week, emerged as the most significant predictors. While some built environment factors such as the presence of certain landmarks and weather conditions were influential, they were less so than temporal cycles. Interestingly the model over-estimates footfall in the years following the COVID-19 pandemic, suggesting that urban dynamics have yet to return to pre-pandemic levels (and may never do). The paper also demonstrates how the model can be used to assess the impacts that large events have had on footfall, which has implications for policy makers as they try to encourage foot traffic back into city centres.