Covariance matrices for crowd behaviour monitoring on the escalator exits

Abstract

The paper describes an approach to detect abnormal events principally from unidirectional flow of crowd (e.g., escalators). The video frames are labeled normal or abnormal based on the distance measure between covariance matrices of the distributions of the optical flow vectors computed on consecutive frames. These flow vectors are the result of tracking a set of features points discovered by the Harris corner detector applied on each frame considering a region of interest. This region is produced by background subtraction to form a two dimensional histogram of motion called motion heat map. The approach is tested against a single camera data-set placed in the escalator exits in an airport. © 2008 Springer Berlin Heidelberg.