Modeling Common Driver Behaviours on Vehicular Data: A Clustering Approach

Simplify and save: companies search for smart fleet management solutions with these two goals that enable them to manage their vehicles easily and efficiently. With this in mind, many tools have been developed, based on several indexes. Among them driving behavior is one of the most common, it is mostly used to evaluate driver safety and to minimize fleet management costs. In this thesis, a highly customizable framework has been proposed. Starting from the data provided by the devices mounted inside the vehicles (Controller Area Network), our tool can determine the driving behavior. To do this, several phases are performed whose outputs are effortlessly interpreted by a person not an expert in data analysis, as a fleet manager. Particular attention was given to data cleaning, modeling and clustering. Supervised machine learning techniques (Random Forest and Gradient Boosting) have been applied to automatically classify vehicles into domain-specific classes. Since we want to group the trips based on the context in which they took place, we analyze the trips travelled on the same route. So knowledge takes on three levels of granularity during the different phases of the pipeline: GPS point, journey and route. The goal is to identify similar trips on the same route, to do this DBSCAN, a consolidated density-based clustering algorithm, is applied. It determines the driving behavior of drivers, highlighting any anomalies. The experiments showed how anomalous trips within a route can be correctly highlighted by the developed software. The outputs given allow clear visualization of which trips are being analyzed and why they are assigned a certain driving behavior.