Can covert and explicit “leaders” steer and split real human crowds?

Poster Presentation: Tuesday, May 21, 2024, 2:45 – 6:45 pm, Pavilion
Session: Action: Locomotor, flow, steering

Kei Yoshida1 (), Hector Taylor1, William H. Warren1; 1Brown University

Collective motion in human crowds is a self-organizing phenomenon that emerges from local visual interactions between neighboring pedestrians. In our previous work, we developed a method of reconstructing spatially-embedded visual influence networks and quantifying local and global influence. We found that some positions in the crowd are more influential than others (“emergent leadership”), as are some participants (“individual leadership”). To test these influential positions, here we experimentally manipulate leadership and network topology in real crowds. In each session, a group of participants (N=16 to 22) was instructed to walk across a field together. One to four confederates, whose presence was either unknown (covert leaders) or visually specified (explicit leaders), turned midway through the trial. In each trial, they received instructions about their initial position (e.g., front or middle of crowd, left/right/center) and turn direction (right, left, no change) via smart watches, and when to initiate the turn via pagers. Head positions were recorded using a Mavic 3 drone with a top-down video camera (60 Hz). To analyze the video recordings, we developed a data processing and analysis pipeline. We trained a convolutional neural network model that uses a multiple-object tracking framework to detect the locations of participants and extract their trajectories (tracklets) and identities. We then used our previous network reconstruction methods to recover visual influence networks and analyze leadership and crowd dynamics. The results show that confederates exerted more influence than non-confederates, particularly when they occupied influential positions. Covert and explicit leaders altered the network topology. We are currently comparing the ability of confederates to steer and split the crowd with predictions from our multiagent simulations of our collective motion model. The results have potential applications to directing emergency evacuations.

Acknowledgements: Supported by NSF BCS-1849446, NIH 1S10OD025181