Members of the defense committee:
- Prof Dr Francesco Viti, University of Luxembourg, Chairman
- Dr German Castignani, University of Luxembourg, Vice-chairman
- Prof. Dr Thomas Engel, University of Luxembourg, Supervisor
- Prof. Dr Marco Fiore, CNR – IEIIT, Italy, Expert
- Dr Johan Wahlström, Oxford University, UK, Expert
- Dr Fabian Lanze, Huf Secure Mobile GmbH, Germany, Expert & advisor
Abstract:
The exponential growth of car generated data, the increased connectivity, and the advances in artificial intelligence (AI), enable novel mobility applications. This dissertation focuses on two use-cases of driving data, namely distraction detection and driver identification (ID).
Low and medium-income countries account for 93% of road traffic deaths; moreover, a major contributing factor to road crashes is distracted driving. Motivated by this, the first part of this thesis explores the possibility of an easy-to-deploy solution to distracted driving detection. Most of the related work uses sophisticated sensors or cameras, which raises privacy concerns and increases the cost. Therefore a machine learning (ML) approach is proposed that only uses signals from the CAN-bus and the inertial measurement unit (IMU). It is then evaluated against a hand-annotated dataset of 13 drivers and delivers reasonable accuracy. This approach is limited in detecting short-term distractions but demonstrates that a viable solution is possible.
In the second part, the focus is on the effective identification of drivers using their driving behavior. The aim is to address the shortcomings of the state-of-the-art methods. First, a driver ID mechanism based on discriminative classifiers is used to find a set of suitable signals and features. It uses five signals from the CAN-bus, with hand-engineered features, which is an improvement from current state-of-the-art that mainly focused on external sensors. The second approach is based on Gaussian mixture models (GMMs), although it uses two signals and fewer features, it shows improved accuracy. In this system, the enrollment of a new driver does not require retraining of the models, which was a limitation in the previous approach. In order to reduce the amount of training data a Triplet network is used to train a deep neural network (DNN) that learns to discriminate drivers. The training of the DNN does not require any driving data from the target set of drivers. The DNN encodes pieces of driving data to an embedding space so that in this space examples of the same driver will appear closer to each other and far from examples of other drivers. This technique reduces the amount of data needed for accurate prediction to under a minute of driving data. These three solutions are validated against a real-world dataset of 57 drivers. Lastly, the possibility of a driver ID system is explored that only uses floating car data (FCD), in particular, GPS data from smartphones. A DNN architecture is then designed that encodes the routes, origin, and destination coordinates as well as various other features computed based on contextual information. The proposed model is then evaluated against a dataset of 678 drivers and shows high accuracy. In a nutshell, this work demonstrates that proper driver ID is achievable even with limited amounts of data. The constraints imposed by the use-case and data availability negatively affect the performance; in such cases, the efficient use of the available data is crucial.