Driver behavior modeling : applications for driving safety assessment and drivers fingerprinting

Abstract

The recent computerization of cars, together with the development of sensor technologies and car communication devices have revolutionized the way researchers and analysts deal with driving behavior. Driving behavior analytics have thus emerged as an important means of improving driving safety and drivers comfort. Depending on the analysis's goals, di erent mathematical and statistical models have been used and numerous analytics approaches have emerged consequently. Generally, these analytics solutions process data generated by vehicles, solely or combined with road data, and transform it into valuable information to gain a better understanding of drivers behavior. In this PhD thesis, we investigate the application of some well-known approaches in driver behavior modeling and analysis. We aim to provide a holistic framework for driver behavior analysis based on automotive sensors data (such as speed, acceleration, steering angle, etc). Our contributions relate to di erent steps involved in the data analysis process, mainly those of preprocessing, modeling and analysis. As a very rst step, to prepare the driving data for analysis, abstraction using the interval domain is proposed. This abstraction transforms the measurements into a form of intervals, ignoring thus irrelevant details from instantaneous measurements. Then, we propose two graphical models to represent driving behavior, which are probabilistic hybrid automata (PRHIOA) and attributed directed graphs (ADG). In fact, probabilistic graphical models provide a helpful framework for modeling driving behavior: the language of graphs facilitates the representation of the relationships within the driver, vehicle, environment system, while the probability allows the representation of uncertainty. The models are combined with a machine-learning algorithm, based on the learning automata algorithm, to build personalized models of the drivers behavior. Besides their representative power, the models proposed have allowed us to perform profound analyses using formal veri - cation and graph matching techniques. Finally, based on these models, two analyses are proposed: the rst analysis uses model checking of the automaton model of the driver to verify the compliance of his driving with the road rules; while the second uses graph matching theory to compute the similarity between the behavior of di erent drivers. Obtained results reveal the potential of the two approaches in improving driver behavior logs analysis.