A machine learning based personalized system for driving state recognition

Abstract Reliable driving state recognition (e.g. normal, drowsy, and aggressive) plays a significant role in improving road safety, driving experience and fuel efficiency. It lays the foundation for a number of advanced functions such as driver safety monitoring systems and adaptive driving assistance systems. In these applications, state recognition accuracy is of paramount importance to guarantee user acceptance. This paper is mainly focused on developing a personalized driving state recognition system by learning from non-intrusive, easily accessible vehicle related measurements and its validation using real-world driving data. Compared to conventional approaches, this paper first highlights the necessities of adopting a personalized system by analysing feature distribution of individual driver’s data and all drivers’ data via advanced data visualization and statistical analysis. If significant differences are identified, a dedicated personalized model is learnt to predict the driver’s driving state. Spearman distance is also drawn to evaluate the differences between individual driver’s data and all drivers’ data in a quantitative manner. In addition, five categories of classifiers are tested and compared to identify a suitable one for classification, where random forest with Bayesian parameter optimization outperforms others and therefore is adopted in this paper. A recently collected dataset from real-world driving experiments is adopted to evaluate the proposed system. Comparative experimental results indicate that the personalized learning system with road information significantly outperforms conventional approaches without considering personalized characteristics or road information, where the overall accuracy increases from 81.3% to 91.6%. It is believed that the newly developed personalized learning system can find a wide range of applications where diverse behaviours exist.