Extracting Microscopic Pedestrian Characteristics from Video Data

Insight into walking behavior is essential for theory and model development describing the behavior of individual pedestrians. In turn, microscopic pedestrian simulation models can be used to test and compare different infrastructure designs, both from the perspective of efficiency and safety. To calibrate these microscopic models, detailed data is required. However, currently such data is not available. This paper discusses an approach to extract individual pedestrian data from digital video footage. In this specific case, the footage was collected during a large scale walking experiment performed at the Delft University of Technology. In this experiment, a group of approximately 80 pedestrians conducted different tests, while being observed using a video camera mounted at the ceiling of the hall. In the first phase of the pedestrian tracking process, different image processing operations, such as radiometric correction and correction for lens distortion are applied. Next, the pedestrians are detected by identifying the special colored caps the pedestrians were wearing during the experiment. This detection occurs in a special zone where conditions are such that pedestrian detection is optimal. Next, pedestrian tracking is achieved by application of a newly developed tracking technique. This is achieved by minimizing the so-called merit-function. A Kalman filter is applied to reduce the errors made during tracking. The paper concludes by showing some of the tracking results, and accuracy estimates. It turns out that the pedestrian trajectories can be determined with high accuracy. Also, some interesting findings pertaining to the behavior of pedestrians are reported.