Exploitation of multi-camera configurations for visual surveillance

In this paper, we propose novel methods for background modeling, occlusion handling and event recognition by using multi-camera configurations. Homography-related positions are utilized to construct a mixture of multivariate Gaussians to generate a background model for each pixel of the reference camera. Occlusion handling is achieved by generation of the top-view via trifocal tensors, as a result of matching over-segmented regions instead of pixels. The resulting graph is segmented into objects after determining the minimum spanning tree of this graph. Tracking of multiview data is obtained by utilizing measurements across the views in case of occlusions. Finally, the resulting trajectories are classified by GM-HMMs, yielding better results for using all different trajectories of the same object together.