Congestion detection in pedestrian crowds using oscillation in motion trajectories

摘要

Identifying crowd congestion in high density pedestrian crowds is a challenging problem, with substantial interest for safety and security applications. We present a method to identify and localize congestion in high density crowd videos, based on the characteristic motion of individuals stuck in congestion. Pedestrians who cannot move freely due to congestion tend to undergo lateral oscillations. Using only motion features, extracted from optical flow and particle advection, we generate sampled trajectories and compute areas which show a pattern of increasing trajectory oscillation. A new and first ever data set including 15 different crowd scenes is introduced to better evaluate the accuracy of our method. Showing results on a diversity of challenging scenarios, we both qualitatively and quantitatively show that this approach provides accurate detection and excellent localization of congestion events.