Detection and tracking people using multiple cameras

One of the most interesting areas of research in computer vision is segmentation and tracking of people using monocular or multi-view systems. In this thesis we present and implement a tracker, which is capable to detect and track people using multiple cameras. Algorithm is incrementaly building a model called mixture of gaussians for each pixel independently. If the current observation does not match its model, then the appropriate pixel is marked as a foreground object (person). From those pixels we create a color representation for each foreground object. Considering color models and probable positions of the people, we track those people across the current scene. To precisely determine the ground location of a person, we map vertical axis of the person (principal axis) to a top-view plane by using homographies. The results show that this approach performs effectively when tracking individual person. However some problems are observed in situations where we monitor several occluded people in a cluttered scene