Crowd Segmentation Method Based on Trajectory Tracking and Prior Knowledge Learning

Abstract

Ensuring public safety and preventing accidents, such as stampedes, caused by crowd congestion, have become urgent problems to be solved. Conflicting, divergent movements in crowds are a main cause of crowd-related safety incidents and accidents. An approach based on trajectory tracking and prior knowledge learning is proposed to handle the problem of crowd motion pattern segmentation. The method uses an orientation–density-function-based cumulative probability model to construct the particle flow field and obtain coherent crowd motion trajectories. Prior knowledge acquisition by trajectory clustering with the help of the improved affinity matrix in the spectral clustering algorithm follows. It then obtains the referential number of clusters automatically and gets the motion pattern segmentation. Experimental results demonstrate that the method presented is effective in identifying the homogeneous crowd motion components under different scenarios. This capability can enable an early warning to be given when a crowd viewed on a video surveillance system is in a congested state.