Roderick Johnstone MSc Dissertation 2015/16

Visual analysis of crowd senes

Supervised by Y.Gotoh

Abstract

People counting and rudimentary behaviour analysis (detection of running) is demonstrated on the PETS2009 dataset for a crowd monitoring method making use solely of sparse feature point positions and velocities. The rationale is the elucidation of the usefulness of motion data alone for crowd monitoring.The method uses optical flow coupled with Kalman filters to obtain motion data, and uses a hierarchical clustering technique based on BIRCH to estimate the number of individuals in a scene. Test results suggest that accuracy degrades for high crowd densities, but may be adequate for detection of high crowd densities, which, coupled with motion information, might suffice for the detection of congestion. The method appears suitable for real-time use.