Tracking Pedestrians Under Occlusion Using Multiple Cameras

This paper presents a integrated solution to track multiple non-rigid objects (pedestrians) in a multiple cameras system with ground-plane trajectory prediction and occlusion modelling. The resulting system is able to maintain the tracking of pedestrians before, during and after occlusion. Pedestrians are detected and segmented using a dynamic background model combined with motion detection and brightness and color distortion analysis. Two levels of tracking have been implemented: the image level tracking and the ground-plane level tracking. Several target cues are used to disambiguate between possible candidates of correspondence in the tracking process: spacial and temporal estimation, color and object height. A simple and robust solution for image occlusion monitoring and grouping management is described. Experiments in tracking multiple pedestrians in a dual camera setup with common field of view are presented.