The recent proliferation of electric scooters, following regulatory changes, has increased traffic accidents, highlighting the need for driver-assistance systems (DAS) that can detect and mitigate hazards. This study aims to develop a DAS that dynamically adapts its support based on the driver's cognitive state. Our prior work attempted to estimate driver awareness using a Support Vector Machine (SVM) model; however, it failed to achieve practical classification accuracy and suffered from poor generalization performance across different subjects. These challenges stem from the inherent difficulty of acquiring large-scale datasets through labor- and time-intensive experiments. To overcome these limitations, this study leverages a time-series foundation model pre-trained on a vast corpus of diverse time-series data. By fine-tuning this model on a small task-specific dataset, we successfully achieve both high classification accuracy and robust generalization performance that is independent of individual driver’s habits. The difficulty of data collection is a common challenge across the broader field of mobility, including other personal mobility vehicles and autonomous vehicles. The insights from this research hold significant potential for application in developing more intelligent and data-efficient safety systems.
Tsubomoto et al. (Wed,) studied this question.