PeTrack

For the understanding of the dynamics inside crowds reliable empirical data are needed enabling an increase of safety and comfort for pedestrians and the design of models reflecting the real dynamics. Manual procedures for collecting this data are very time-consuming and usually do not supply sufficient accuracy in space and time.

For this reason we are developing the tool named PeTrack (Pedestrian Tracking) to automatically extract accurate pedestrian trajectories from video recordings. The joint trajectories of all pedestrians provide data like velocity, flow and density at any time and position. With such a tool extensive experimental series with a large number of persons can be analyzed. Individual codes enables personalized trajectories with static information of each participant (e.g. age, gender).

The program has to deal with wide angle lenses and a high density of pedestrians. Lens distortion and perspective view are taken into account. The procedure includes calibration, recognition, tracking and height detection.

Different kinds of markers (e.g. with height information, head direction, individual code) are implemented. With a stereo camera more accurate height measurements and also markerless tracking is possible. Machine learning methods also enable markerless detection of people. In addition, shoulder orientation can be captured, comments and metadata can be annotated, fused motion capture data can be visualized, and corrections can be made to the trajectories.

The source code as well as precompiled executables of PeTrack are available. For developer you find information how to contribute to the code base here, like setting up an IDE, code conventions or how to use the testing framework.

This video gives an overview of steps for extracting trajectories.

The comprehensive documentation also includes a starting guide with an exemplary project and how to prepare experiments for high quality data. The documentation of using PeTrack cannot answer all questions. Thus you are also welcome to contact the authors before setting up experiments and an automatic extraction with PeTrack: petrack@fz-juelich.de.

Results of collected trajectories can be found at our data archive of studies about pedestrian dynamics.