A GM-PHD filter is used for pedestrian tracking in a crowd surveillance application. The purpose is to keep track of the different groups over time as well as to represent the shape of the groups and the number of people within the groups. Input data to the GM-PHD filter are detections using a state of the art algorithm applied to video frames from the PETS 2012 benchmark data. In a first step, the detections in the frames are converted from image coordinates to world coordinates. This implies that groups can be defined in physical units in terms of distance in meters and speed differences in meters per second. The GM-PHD filter is a Bayesian framework that does not form tracks of individuals. Its output is well suited for clustering of individuals into groups. The results demonstrate that the GM-PHD filter has the capability of estimating the correct number of groups with an accurate representation of their sizes and shapes.