Automatische Erfassung präziser Trajektorien in Personenströmen hoher Dichte

Simulations can help make facilities for pedestrians safer and more comfortable. A proper understanding
of crowd dynamics is essential to developing reliable models for such simulations. Detailed and
reproducible datasets of real crowd movements are needed for analysis and modelling. Such datasets
are also required for later calibration and validation of said models.
This thesis describes the collection of such data from overhead video recordings. Individual trajectories
are extracted and make it possible to obtain the most relevant quantities of the dynamic e.g.
pedestrian density, velocity and flow. Traffic jams and other high density situations are of special interest
since this is where critical situations are to be expected. Therefor the developed methods have
to also reliable extract an individual’s movement in such situations.
The movement of pedestrians is affected by many factors such as geometry, crowd density, motivation
and culture. To investigate these numerous influences a large number of experiments with a huge
number of participants have been carried out. The automatic extraction of the trajectories provides
a significant advantage compared to manual methods in terms of the time required, accuracy and
reproducibility.
The extraction process consists of the image calibration followed by the detection, tracking and determination
of the real world position of all individuals. For the detection of a person various markers and
corresponding extraction techniques have been developed for the different applications and local conditions.
A markerless method was also developed, which is especially useful for field studies. Through
the use of stereo cameras high detection rates were achieved without markers, even in high density
situations. All developments regarding the extraction process have been integrated into the software
PeTrack. To enable a deeper understanding of the results the technical aspects of the trajectory
collection are described in addition to the recognition techniques.