Advancements in intelligent driving assistance: A machine learning approach to identify real-time driving strategies using environmental, eye movement, control-related, and kinetic-related data

Abstract

Human-vehicle collaboration (HVC) is gaining focus as intelligent vehicle technology shifts driving responsibilities from humans to automation. Effective HVC hinges on mutual understanding between parties. Driving strategies address specific driving requirements and guide maneuver decisions, offering a gateway to uncovering a driver’s immediate needs. The present study introduces the framework of adaptive driving (FAD), which uses machine learning algorithms to incorporate critical features for identifying four common driving strategies, including competitive driving, skill-based behavior, defensive driving, and urgent reactions. A total of 52 participants engaged in simulated driving sessions, with these sessions yielding 983 driving events that were used for model training. Discriminative models considering driving environment, eye movement, or combined features can satisfactorily classify four driving strategies, while control-related and kinetic-related data offer minor contributions. Environmental data significantly aids in identifying skill-based and competitive strategies, characterized by lower risk dynamics and longer Time-to-Collision (TTC), with forward targets often triggering competitive driving. For risk mitigation strategies, both environment and eye movement data are crucial. Scenarios with higher risk dynamics prompt defensive driving, while shorter TTC indicates urgent reactions. Defensive driving is revealed by saccade patterns and less variation in pupil activities, whereas urgent reactions show pronounced pupil dilation and reduced saccade activity, indicating challenges in environmental observation during emergencies.