Visibility management for object tracking in the context of a fisheye camera network

Fisheye lenses provide a wide field of view of the scene that can help reduce the number of cameras used to cover a large area. However, the geometric deformations induced by such devices can make the image interpretation difficult. For this reason, the presence of many obstacles, which is a difficult problem in every computer vision system, is especially impairing in a fisheye context, notably in the peripheral areas of the image where object visibility is already low. In this paper, we describe several generic strategies to deal with partially visible objects in a highly cluttered scene. We explain how a global model of the static obstacles can help in dealing with such problems. We present an implementation of our ideas in a real-time system backed by a network of fisheye cameras tracking customers in a retail store.