What epipolar geometry can do for video-surveillance

In this paper we deal with the problem of matching moving objects between multiple views using geometrical constraints. We con- sider systems of still, uncalibrated and partially overlapped cameras and design a method able to automatically learn the epipolar geometry of the scene. The matching step is based on a functional that computes the similarity between objects pairs jointly considering different contri- butions from the geometry. We obtain an efficient method for multi-view matching based on simple geometric tools, requiring a very limited hu- man intervention, and characterized by a low computational load. We will discuss the potential of our approach for video-surveillance applications on real data, showing very good results. Also, we provide an example of application to the consistent labeling problem for multi-camera tracking, and report a comparative analysis with other methods from the state of the art on the PETS 2009 benchmark dataset.