Revisiting Gaussian Mixture Models for Driver Identification

Abstract

The increasing penetration of connected vehicles nowadays has enabled driving data collection at a very large scale. Many telematics applications have been also enabled from the analysis of those datasets and the usage of Machine Learning techniques, including driving behavior analysis, predictive maintenance of vehicles, modeling of vehicle health and vehicle component usage, among others. In particular, being able to identify the individual behind the steering wheel has many application fields. In the insurance or car-rental market, the fact that more than one driver make use of the vehicle generally triggers extra fees for the contract holder. Moreover being able to identify different drivers enables the automation of comfort settings or personalization of advanced driver assistance (ADAS) technologies. In this paper, we propose a driver identification algorithm based on Gaussian Mixture Models (GMM). We show that only using features extracted from the gas pedal position and steering wheel angle signals we are able to achieve near 100% accuracy in scenarios with up to 67 drivers. In comparison to the state-of-the-art, our proposed methodology has lower complexity, superior accuracy and offers scalability to a larger number of drivers.