Predicting Driving State, Intent to Merge and Working Memory Load from Eye-Tracking Data Collected in a Driving Simulator Study

Eye tracking has gained renewed attention with camera-based systems and machine-learning methods that enable non-invasive inference of cognitive state. This thesis explores whether eye-tracking features can predict (i) driving state, (ii) intent to merge, and (iii) working-memory load in a virtual driving environment. Participants drove through unsignaled intersections with oncoming traffic while performing an auditory n-back task that elicited low or high working-memory load.We demonstrate robust cross-subject classification of driving state and significant classifier performance for most participants for both intent to merge and working-memory load. Methodologically, we introduce a bivariate Gaussian density feature of horizontal and vertical gaze that improves driving-state and intent-to-merge classification. We also show that pupillometric features—especially differences in pupil-size variability between fixations and saccades—aid working-memory classification in this scenario, and we confirm the utility of median/mean pupil size reported in prior experimental work.Overall, our findings support growing evidence that eye-tracking signals can inform an operator’s cognitive state and may translate to real-world assistance systems. However, because a virtual environment cannot capture all real-world confounds, these results should be viewed as evidence of potential rather than proof of real-world utility.