Abstract
A bottom-up approach to understanding crowd behavior begins by modeling local interactions between pairs of pedestrians, and then scaling up to simulate the collective motion of crowds. Previously, we developed a dynamical model of pedestrian following, and we now test whether the model can simulate empirical data on collective motion. In earlier experiments, we found that a follower matches the leader's speed (Rio, Rhea, & Warren, 2014) and heading direction (Dachner & Warren, 2014); we also determined that the influence of neighbors in a crowd decreases linearly with distance (Warren & Rio, VSS 2015). Here we test how well the model simulates data previously collected in the Sayles Swarm (Warren & Bonneaud, VSS 2014) on groups of pedestrians walking together. Head position was recorded with 16 motion-capture cameras in a 12m x 20m tracking area. First, we model groups of 4 pedestrians walking 20m (initial interpersonal distance 1, 2, 4 m). On each trial, participants were instructed to turn twice (left-right; right-left; left-left; or right-right) or change speed twice (slow-fast; fast-slow; slow-slow; or fast-fast) as a group, at a self-selected time. Second, we model groups of 20 participants who were instructed to walk about the tracking area for 2 min, veering randomly left and right while staying together as a group (initial interpersonal distance 1, 2 m). We simulate one pedestrian at a time, taking the data from other visible pedestrians as input, and generate a simulated trajectory as output, for each trial. The time series of the pedestrian's speed and heading are compared to those of the simulated agent, with R2 and RMS error as dependent measures. Initial results demonstrate a good fit to the human data, with fixed parameters. An empirically-based following model thus successfully scales up to simulate the collective motion of pedestrian groups.
Meeting abstract presented at VSS 2017