Model-Based Occlusion Handling for Tracking in Crowded Scenes

The task of reliable detection and tracking of multiple objects becomes highly complex for crowded scenarios. Data association is difficult to perform reliably in the presence of missing observations due to occlusions. We propose a novel real-time approach to segment and track multiple overlapping humans. The optimal segmentation solution is given by the maximum likelihood estimate in the joint-object space. The search for solution is guided by a fast mean shift procedure and relies on information on the number of humans involved in the occlusion which can be estimated using the tracking history. Results are presented for the task of human tracking in crowded scenes and evaluated in terms of tracking performance.