Abstract
Walking on our daily commutes, we encounter dozens of other pedestrians but rarely do we collide with them. Previous omniscient models of collision avoidance assume 3D positions and velocities are known and predict the trajectories of all obstacles, which is not biologically plausible. Bai & Warren (VSS 2021, 2022) developed a visual model that accurately predicts human collision avoidance of a single moving obstacle, based on change in its bearing direction and visual angle. Here we test the model on avoiding single or multiple walking pedestrians. Participants walked to a goal at 7m while avoiding moving avatars presented in a head-mounted display. Experiment 1 (N=12) compared avoidance of a moving pole and a walking avatar. The obstacle moved on a linear path at different angles (180°, ±157.5°, ±135°, ±112.5°, ±90° to the participant’s path) and speeds (1.0, 1.2 m/s), and the participant’s head position was recorded. There was no difference in trajectories around a pole and an avatar (BF10 = 0.22), which the model predicted equally well (Mean RMSE=0.173m, SD =0.38m and Mean RMSE=0.178m, SD=0.38m, respectively). The model thus generalizes from avoiding inanimate obstacles to avoiding walking pedestrians. In Experiment 2 (N=12), participants walked through a crowd of 16 avatars, which moved at the same speed (1.0 m/s) in a common direction (same angles as before). We compare the performance of models that avoid avatars closer than 4m (Mean RMSE=0.698m, SD =0.48m) or avoid avatars that exceed a visual threshold for optical expansion and bearing change (Mean RMSE=1.01m, SD =0.483m), and are investigating other variations of the visual model. Together, the results demonstrate that a visual model of collision avoidance with a single moving obstacle generalizes to the avoidance of multiple collisions when walking through a crowd.