Abstract
The global behavior of human crowds is thought to emerge from local, visually-guided interactions between pedestrians. There are numerous physical and computer models of such collective behavior (Helbing & Molnar, 1995; Moussaid, et al., 2010; Pettré, et al, 2009), but few are based on experimental evidence. We are building an empirical "pedestrian model" of locomotor behavior and using agent-based simulations to test whether local interactions are sufficient to generate characteristic patterns of crowd behavior. We have derived model components for steering to a goal, avoiding obstacles, matching the speed and heading of neighbors, and braking. Here we report new data on human crowds and our attempt to model them using a multi-agent simulation platform. We collected naturalistic data on groups of up to 20 participants walking in several scenarios in an open area (12x20 m). Head positions were tracked using a 16-camera motion capture system (Qualisys, 60 Hz). We present videos of the following scenarios, corresponding head trajectories, and animations of qualitative simulations: (1) In the "Grand Central Station" scenario, 20 participants criss-cross an arena for 3 min, while avoiding 10 obstacles and each other. (2) In the "swarm" scenario, 20 participants randomly veer left and right while staying together as a group for 2 min. (3) In the "counterflow" scenario, two groups pass through each other, yielding spontaneous lane formation. In each case, the global pattern is reproduced with a few model components. The results provide evidence that the global behavior of human crowds is emergent and may be accounted for by local pedestrian interactions. We are pursuing questions such as: What is the strength of the visual coupling between neighbors? Can the model simulate individual trajectories? Is swarming a self-organized behavior? Do coherent swarms form spontaneously or derive from explicit switching between control laws?
Meeting abstract presented at VSS 2013