Youjin Oh¹, Nathaniel Powell¹, Daniel Panfili¹, Mary Hayhoe¹; ¹University of Texas at Austin

In complex natural environments it is unclear how attention is directed to information needed to avoid independently moving obstacles. To examine behavior in such contexts we recorded gaze, head, and foot movements and 3D scene data while subjects walked through crowded sidewalks. Gaze was recorded using a Pupil Labs Core mobile eye tracker which was integrated with a head-mounted Intel RealSense 435i stereo camera used to extract depth relative to gaze location. Objects within the video feed of the RealSense depth camera were classified using a YoloV3 object detection algorithm. The results provide a description of the gaze behavior people exhibit in the presence of moving objects like other pedestrians, and how it varies based on distance from the walker. Preliminary data show frequent gaze shifts to distant objects and regions of space. Pedestrians were fixated when they were around 5 to 8 meters away from the walker, depending on pedestrian density. This distance corresponds to roughly 10 footsteps away, suggesting that walkers adopt a proactive policy for detecting potential hazards. In the context of this strategy, the total time looking at pedestrians was quite short, even with high pedestrian density. About one-tenth of the total time was spent directly fixating other pedestrians. Thus only short time periods of direct gaze on potential obstacles are necessary, and walkers spend most time looking in the direction of the future walking path. The effectiveness of this strategy indicates that the predictability of pedestrian behavior allows efficient use of gaze in these contexts. By remaining sensitive to the statistics of the natural world, subjects can allocate attention efficiently and avoid reliance on bottom-up mechanisms (Jovancevic et al. 2009). Our results add to the understanding of how the visual system extracts the necessary information to adapt to the dynamic environment.