Predicting natural road scenes: Effects of motion, stereoscopic depth, and scene context

Driving is an inherently visual task that requires drivers to anticipate and predict future events on the road. We investigated the extent to which three different visual features – motion, stereoscopic depth, and scene context—can support drivers’ prediction of the appearance of natural road scenes. On each trial, 48 licensed drivers viewed a 2s preview of a video recorded from a dashboard camera, and immediately after, indicated what the scene would look like 2s in the future in a 5AFC task. We factorially manipulated each of the three visual features. We manipulated motion by displaying a still image or a video preview. We manipulated scene context by showing videos recorded either on urban roads or on highways. To investigate the impact of stereoscopic depth information, videos were recorded using two dashboard cameras, and videos were shown either binocularly with no disparity, binocularly with disparity, or monocularly. Overall, most participants were able to predict the appearance of the road above chance levels, with better performance for video previews than still images, and for urban roads than highway roads, indicating that speed and scene content inform predictions. Stereoscopic cues did not affect prediction performance, indicating that participants likely rely more heavily on monocular depth cues in this task. These results suggest that participants can imprecisely predict natural scenes multiple seconds into the future.