Tracking People in Crowded Scenes

Abstract

For the collection of trajectory data of pedestrianmovement in crowdswe are developing software for the automatic extraction of accurate trajectories out of video recordings. In this paper a newly introduced method for a markerless detection is presented. This method identifies people by comparing the shape of the top part of their body with the perspective depth field of overhead stereo recordings. Features to match are pyramidal stacks of ellipses, which approximate isolines of the same distance to the camera.