Data-driven simulation of pedestrian dynamics: collision avoidance between two moving agents

MARTIN, RAFAEL F.; PARISI, DANIEL R.

Melbourne (y Virtual por pandemia)

Conferencia; Pedestrian and Evacuation Dynamics 2021; 2021

In recent works, we proposed a general methodology for a data-driven pedestrian simulation that uses neural networks as a way to access and interpolate the experimental data. This methodology avoids the use of mathematical and physical models and allows simulations to be generated directly from experimental data. This has the following advantages: - avoid searching for intrinsic parameters of the models that may or may not correspond to reality; and provide, by definition, experimentally validated simulations. As a first case, it was used to simulate the simple problem of a pedestrian avoiding a fixed obstacle. In the present work, we will address, with the same methodology, the problem of avoiding collisions between two moving pedestrians. The necessary data were recorded in our motion capture laboratory that provides trajectories with high temporal and spatial resolution. Because of this, the amount of data can be excessive to describe a given trajectory, so we study which is the density of data that optimizes the performance of the method. The proposed methodology, together with the database of real trajectories, has produced satisfactory results that allow simulating trajectories of two pedestrians with potential collisions from different angles. Finally, it is proposed how to extend this simulation scheme to low-density scenarios with more than two pedestrians in motion.