Modeling Self-Organization in Human Crowds
Abstract
Developing an understanding of human crowd movement is critical for optimizing crowd safety,
both on the streets and in buildings. It is well understood from the study of animal collective motion
that individuals are better able coordinate and respond to perturbations by perceiving movements
from their neighbors (Ballerini et al., 2008; Couzin & Krause, 2003; Partridge, 1982). Thus,
modeling and simulating human crowd behavior depends on the modeler’s ability to accurately
characterize how an individual is influenced by his or her neighbors (Ballerini et al., 2008); i.e.,
identifying the local rules of engagement that lead to self-organization (Rio et al., 2018) – a process
in which coordination emerges solely from the interactions of these local rules (Couzin & Krause,
2003).
Popular models that are used today for human crowd simulation are theory-driven (Moussaïd et
al., 2009, 2011; Sumpter et al., 2012), based on particle behavior (Helbing & Molnar, 1995) as
opposed to data-driven, based on human behavior (Rio et al., 2018). This creates a serious problem:
These models generate local trajectories that do not realistically simulate pedestrian behavior,
resulting in many collisions between individuals (Pelechano et al., 2007). We at the VENLab
employ an alternative approach: develop a data-driven crowd model by studying real human
behavior.
Previous VENLab experiments have demonstrated that when walking with a crowd, pedestrians
form their alignment decision by averaging over the local neighborhood within a soft-metric radius
of about 5m (Rio et al., 2018; Wirth, Dachner, Rio, & Warren, in Review; Willcoxon & Warren,
in Preparation). In further exploring the rules of engagement that lead to self-organized crowd
behavior, there are two critical questions that my dissertation work addresses: 1. How is an
individual pedestrian recruited into collective motion? And 2. When a pedestrian is following a
crowd that splits in disparate directions, what rules dictate which group the pedestrian follows?
While the dissertation consists of five empirical studies – with model modifications that were
inspired by the experimental insights – I will report the two most important findings here that
answer the questions posed above. An unfinished study, which is currently being completed, is
discussed in Future Directions.