BIOND LAB

The movement of pedestrian crowds is studied both for real-world applications and to gain fundamental scientific insights into systems of self-driven particles. Trajectory data describes the dynamics of pedestrian crowds at the level of individual movement paths. Analysing such data is a central challenge in pedestrian dynamics research, coupled with increasing data availability this implies a need for efficient methods to identify key features of the captured crowd dynamics. In this paper, we show that diffusion maps, an unsupervised manifold learning method, can be used for this purpose. We show how to build an informative feature space by defining a set of observables from trajectories. We use our diffusion map approach to analyse pedestrian movement on a stadium-shaped track, and during egress from a room, considering hundreds of trajectories for each scenario. We first verify that our diffusion map analysis can recover known leading variables that determine the system dynamics. Then, we show how our analysis facilitates a qualitative comparison of the dynamics inherent in entire data sets, by contrasting experimental with numerically simulated data. Finally, we establish how our approach can be used to automatically detect outliers that show behaviour distinct to others. These results indicate that our work can contribute a computationally efficient and unsupervised approach to analyse pedestrian dynamics without needing much prior knowledge of the data. We suggest this could be useful for automatically monitoring live data, or as an initial step to inform a subsequent analysis.