Light Footprint Driving Behaviour Classification

Automatic classification of driving behaviour from data collected by vehicle sensors is a widely studied problem in the intelligent transportation community. Most papers on this subject assume that, when classification is performed for a driving segment, complete data from all sensors at their maximum frequency is available for that segment. However, if classification is not performed in the vehicle, such data need to be transferred to a server and stored for later use, leading to significant costs if several millions of vehicles are involved, as for fleet management services. For this reason, in this paper we consider the problem of classifying driver behaviour minimizing the amount of information used (light footprint driving behaviour classification). We propose a machine learning-based feature selection approach that allows us to retain small sets of uncorrelated features with maximal importance in the classification task. We test our methodology on the UAH DriveSet, proving its ability in obtaining good classification results using as few as three features.