Jiali Song¹ (), Benjamin Wolfe¹; ¹Unersity of Toronto Mississauga

Safely responding to a road hazard requires knowing when and where it is in the scene. Conventional eye-tracking studies in traffic safety use direct fixation as a proxy for awareness of the hazard. However, fixation location is a poor indicator of awareness because peripheral vision enables drivers to gather information without direct fixation. Conversely, fixating an object does not necessarily guarantee awareness (e.g. inattentional blindness, looked-but-failed-to-see errors). Given that road scenes change dynamically over time, little is known about when and where drivers look when searching for hazards, particularly when they miss hazards. We investigated these questions by asking drivers (n=8) to watch 262 dashcam videos of near-collisions from the Road Hazard Stimulus dataset. Each video (2-8 sec) contained a hazard that was eventually involved in a near-collision in the left or right half of the video, and participants reported its location. Hazard onset was annotated as the first frame in the video that indicates that a near-collision would occur. We grouped the trials based on whether participants correctly reported the hazard’s location. During the time window immediately prior to hazard onset, drivers’ gaze distance from the hazard decreased over time on correctly localized trials. On incorrect trials, gaze position was significantly further away for 185ms after hazard onset compared to correct trials (t(7)>2.38, p <0.05). Furthermore, participants were more likely to look within the annotated hazard region for correct trials than for incorrect trials, particularly from 480ms before hazard onset and onwards (t(7)>2.37, p<0.05), and intermittently in the 1000ms leading up to the time of response (t(7)>2.37, p<0.05). These results suggest that, on average, drivers look at hazards when they correctly detect them and that looked-but-failed-to-see errors are relatively rare. Rather, incorrect trials are more likely to be “looked too late” or “failed to look” errors.

Acknowledgements: We thank Dr. Anna Kosovicheva for her guidance and advice throughout the design and analysis of this project. This work was funded by a National Science and Engineering Research Council grant awarded to Benjamin Wolfe (RGPIN-2021-02730).