Jiali Song¹, Benjamin Wolfe¹; ¹Unersity of Toronto Mississauga

Prediction is a fundamental visual process and is essential for safe driving. Successfully avoiding a moose on the road requires knowing where it is and where it’s going. Several theories of road hazard perception propose that avoiding collisions requires accurately predicting them several seconds in advance. We tested this claim and examined the extent of agreement among drivers. Twenty-four licensed drivers (mean age 19.7 years, and >1 year of experience) viewed 360ms excerpts from 174 dashcam videos that each contained a collision or near-collision. Four video excerpts were created from each video to manipulate excerpt time relative to hazard onset (moment of the first visible indication of an impending collision). Excerpt time was manipulated in four sequential blocks (negative times represent time before hazard onset): -2000ms, -500ms, 0ms, and just before the driver in the video responded. After viewing each excerpt, participants were asked to report the location of the most likely hazard in the scene by clicking in an onscreen rectangle where the video had been shown. Prediction accuracy was defined as the proportion of clicks that corresponded with the location of the collision. Prediction accuracy increased monotonically with later excerpt time, with the largest increase between -2000ms and -500ms, suggesting that this temporal window is key for accurate predictions. We indexed agreement for each excerpt using average pair-wise distance between all clicks, which decreased monotonically, with the largest decrease between -2000ms and -500ms. Click locations were moderately correlated with fixation locations collected previously in a separate sample of 30 drivers who watched the same videos (mean Pearson’s r = 0.52), suggesting that drivers look at locations that are predicted to contain hazards. These results demonstrate that on-road predictions rely on information that appears 2000ms - 500ms before hazard onset, are moderately consistent across drivers, and guide search strategy.

Acknowledgements: This work was funded by a National Science and Engineering Research Council grant awarded to Benjamin Wolfe.