Event Detection in Crowds of People by Integrating Chaos and Lagrangian Particle Dynamics

This paper proposes a system for automatic video analysis to detect events in video sequences with crowds of people. In detail, the proposed system consists of three subsystems: 1) the first identifies the motion areas, resorting to chaos theory using joint histogram between consecutive frames, 2) the second one creates a flow motion map that describes the behavior of motion pixels by using Lagrangian Particle Dynamics Theory and, 3) the last one uses self organizing maps (SOM) for segmenting the flow motion map in order to detect events. The proposed method was tested on a set of 30 videos, describing crowds in different scenarios, collected from the BBC Motion Gallery, achieving an average accuracy of about 87% in detecting events such as people stopping, people laying on the ground, group of people fighting, obstacles on the road and long queues.