Understanding crowd flow patterns using active-Langevin model

Abstract

Crowd flow describes the elementary group behavior. Dynamics behind group behavior can help to identify abnormalities in flows. Quantifying flow dynamics can be challenging. In this paper, an algorithm has been proposed to describe groups’ movements in crowded scenarios by analyzing videos. A force model has been proposed based on the active Langevin equation, where the motion points are assumed to behave similarly to active colloidal particles in fluids. The force model is further augmented with computer-vision techniques to segment linear and non-linear flows. The evaluation of the proposed spatio-temporal flow segmentation scheme has been carried out with public datasets. Experiments reveal that the proposed system can segment the flows with lesser errors than existing methods. The segmentation accuracy and Normalized Mutual Information (NMI) have improved by 10% as compared to existing flow segmentation algorithms.