Analysis of real surveillance video footage is very challenging - hence this thesis provides solutions to enhance the quality of trajectories of dense crowded scenes in real-time. An efficient algorithm models dense crowded scenes with the aid of particles, moved by the optical flow calculated between two consecutive frames. Thus trajectories are obtained without using a people tracking algorithm.
Sources and sinks are modeled by clustering of start and end points. As dense crowded scenes are analyzed, many trajectories are interrupted thus making the choice of an appropriate clustering algorithm challenging - this thesis provides approaches to enhance the quality of trajectories. Furthermore, it evaluates different clustering algorithms and their practicability in combination with the real-time particle advection algorithm on benchmark data of a Viennese train station and additional data provided by the PETS workshop and the University of Central Florida.