August 2014
Volume 14, Issue 10
Free
Vision Sciences Society Annual Meeting Abstract  |   August 2014
Multi-Agent Simulation of Collective Behavior in Human Crowds
Author Affiliations
  • William Warren
    Dept. of Cognitive, Linguistic and Psychological Sciences, Brown University
  • Stéphane Bonneaud
    Dept. of Cognitive, Linguistic and Psychological Sciences, Brown University
Journal of Vision August 2014, Vol.14, 4. doi:10.1167/14.10.4
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to Subscribers Only
      Sign In or Create an Account ×
    • Get Citation

      William Warren, Stéphane Bonneaud; Multi-Agent Simulation of Collective Behavior in Human Crowds. Journal of Vision 2014;14(10):4. doi: 10.1167/14.10.4.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

The collective behavior of human crowds is thought to emerge from local interactions between pedestrians. There are many models of such "flocking" behavior in animals and humans, but few of them are based on experimental evidence about these visually-guided interactions. We have built an empirically-grounded pedestrian model, including components for steering, obstacle avoidance, matching the speed and heading of neighbors, and braking, expressed as dynamical systems. We test whether the model can generate characteristic patterns of crowd behavior using multi-agent simulations. Here we aim to quantitatively evaluate these simulations against human crowd data. Locally, we compare individual human and model trajectories, which faces an n-body problem (e.g. classical mechanics). Globally, we compare distributions of trajectories (e.g. statistical mechanics). Method. We simulated naturalistic data collected from groups of 20 participants walking in a 12x20m arena (Warren, et al., VSS 2013). Head positions were tracked using a 16-camera motion capture system (60 Hz). In the "Grand Central" scenario, participants criss-crossed the arena, while avoiding 10 obstacles and each other. In the "swarm" scenario, participants veered left and right while staying together as a group. Each trial was simulated by initializing agents with each participant's starting position, heading, and speed, and assigning their final position as the goal; different combinations of model components were tested. Results. A local "divergence" measure computed mean agent-participant distance (error) at each time step as a function of elapsed time. Surprisingly, the mean error over all agents is only 30 cm, with a maximum of 1 m, for the best combination of model components. Measures of global traffic patterns, including occupancy heat maps and route histograms, are also compared. The results suggest that the collective behavior of human crowds can be accounted for by an empirical model of pedestrian behavior and emerges from local interactions.

Meeting abstract presented at VSS 2014

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×