How does the built environment affect driver behaviours? A methodological framework for large-scale analysis that combines Onboard Diagnostic and Geographical Data to promote Road Safety

• In this paper, Onboard Diagnostic and Geographic Data are used to study driver behaviours. • Multi-layer perceptron Neural Networks are implemented to classify driver behaviours. • Feature selection methods are implemented to highlight the contributory factors. • The presence of schools, hospitals, bridges, parking, subways, cycleways, and footways leads to aggressive driver behaviour. • This methodology is suitable for large-scale analysis to improve road safety. Driving behaviours are a set of actions made by drivers during their driving tasks and could be detected with the help of onboard sensors in vehicles. The interest from researchers in evaluating fuel consumption and enhancing overall safety has increased in recent years, thanks to the availability of algorithms and methods that can manage the large amount of data obtained from sensors. In this paper, both Onboard Diagnostic (OBD) data from the DDD20 dataset and open-source Geographic Data are used to study driver behaviours. After the preprocessing phases, a Multi-Layer Perceptron (MLP) classifier is used to classify driver behaviours, and a Shapley Additive Explanations (SHAP) Analysis is implemented to perform feature selection. Three models are created: the first combines OBD and geographical data, the second is based only on geographical data, and the last one contains a reduced subset of features retrieved from geographical data. The model with a reduced number of features shows good prediction accuracy, comparable with the full previous models. In addition to that, SHAP analysis highlights how the presence of schools, hospitals, bridges, parking, subways, cycleways, and footways increases the likelihood of having aggressive driver behaviour. This study aims to show how the external context influences driver behaviours and to create a methodological framework for future developments in road safety. The model that uses only open-source geographical data with a reduced number of features is particularly suited for large-scale analysis in the context of road safety.