The increasing availability of sensor data in cities has resulted in the accumulation of large datasets. offering insight into urban phenomena. In particular, monitoring and analysis of pedestrian traffic serve as valuable assets for optimizing city processes and improving the living conditions and attractiveness of a city. Forecasting pedestrian traffic is a key challenge for efficient resource allocation and emergency response planning. However, equation-based modeling of traffic is difficult because of the inherent complexity and many external factors.
In this thesis, we address these challenges by employing multiple machine learning models as a means of forecasting pedestrian traffic data.
We set a particular focus on neural network models due to their demonstrated success in many data-centered problems. Additionally, we investigate classical autoregressive integrated moving average (ARIMA) models, which are widely used in time series analysis and forecasting. Furthermore, we examine a novel Koopman operator-based approach, that allows for analyzing the dynamics of the underlying system.
Our methodology involves a comparative analysis of these approaches, where we first evaluate their performance on real sensor data from the city of Melbourne. We also highlight the inherent differences and respective advantages of each model.
We find that both the neural network models and the Koopman operator approach achieve excellent performance on this dataset, while the ARIMA models do not adequately explain the data.