ABSTRACT
Traffic flow congestion is a very present problem on the daily life of citizens of big cities. Furthermore, it is growing by the day because of the increase of population. Furthermore, it has undesirable consequences such as an increase of air pollution levels and a worse life quality. Traditional solutions, such as investing on public transport, are less effective nowadays because of the COVID-19 pandemic. A good alternative are traffic flow optimization methods, e.g., signal on-off times optimization methods. However, these methods use traffic simulators that are very time consuming and typically act as a bottleneck for the optimization algorithm. In this work, we study if and how Deep Learning models could replace traffic simulators for a more performant alternative for its use on optimization methods. We design several network architectures and use them to predict vehicle and pedestrian time lost in a specific intersection of the city of Salamanca (Spain). The best of our models has an average Mean Absolute Error (MAE) lower than a second using 10-fold cross-validation. Finally, we discuss mechanisms to generalize our models to other intersections using only a reduced amount of data.