Probabilistic dynamics of small groups in crowd flows

Pedestrians in crowds frequently move as part of small groups, which can constitute up to 70% of individuals in public spaces. Dyads (groups of two) are most frequent. Understanding quantitatively the dynamics of dyads walking in crowds is therefore an essential building block towards a fundamental comprehension of the crowd behavior as a whole, and is mandatory for accurate crowd dynamics models. Unavoidably, due to the non-deterministic behavior of pedestrians, characterizations of the dynamics must be probabilistic. In this work, we analyse the dynamics of over 6 M dyads: a statistical ensemble of unprecedented resolution within a multi-year real-life pedestrian trajectory measurement campaign (about 21 M trajectories, collected at Eindhoven Central Station, The Netherlands). We provide phenomenological models for dyad behavior depending on the surrounding crowd state. We present a thorough collection of fundamental diagrams that probabilistically relate both dyad velocity and dyad formation to the state of the surrounding crowd (density, relative velocity). Depending on the surrounding crowd, dyads adjust their interpersonal distance and may shift in formation, possibly moving from abreast states (known to favor social interaction) to in-file (which favors navigation through dense crowds). To quantitatively investigate formation changes, we introduce a scalar indicator, which we dub Orientation Log-Odds (OLO), that quantifies the relative log-likelihood of abreast versus in-file formations. Conceptually, for any given crowd state, the OLO quantifies the energy difference between the abreast and the in-file configuration under a Boltzmann-like assumption. We model how OLO depends on the crowd state, showcasing that its derivative is a product of two velocity-density fundamental diagrams. Together, these results provide a statistically robust, data-driven description of dyad configuration dynamics in real-world crowds, establishing a foundation towards new predictive, group-aware crowd models.