Assessing data imbalance correction methods and gaze entropy for collision prediction

Abstract

Driver Readiness (DR) refers to the likelihood of drivers successfully recovering control from automated driving and is correlated with collision avoidance. When designing Driver Monitoring Systems (DMS) it is useful to understand how driver states and DR interact, through predictive modelling of collision probability. However, collisions are rare and generate imbalanced datasets. Whilst rebalancing can improve model stability, reliability of correction methods remains untested in automotive research. Furthermore, it is not yet clear the extent to which certain features of driver state are associated with the probability of a collision during critical scenarios. The current study therefore had two general aims. The first was to examine statistical model reliability when using imbalance-corrected datasets; the second was to investigate the predictive utility of gaze entropy and pupil diameter in assessing collision risk during critical transitions of control from a simulated hands-off SAE L2 driving experiment. Dataset rebalancing reduced prediction accuracy and overestimated collision probabilities, aligning with prior findings on its limitations. Erratic, spatially distributed gaze fixations were associated with higher collision probability, whilst increased mental workload (indexed via mean pupil diameter) had minimal impacts. We discuss why in many situations researchers should be wary of rebalancing their datasets, and underscore gaze behaviour’s importance in DR estimation and the challenges of dataset rebalancing for predictive DR modelling.