Model-based initialisation of vehicle tracking: dependency on illumination

Although a model-based vehicle tracking approach offers the promise to be more reliable than a purely data-driven one, based on the additional knowledge brought to bear during the tracking phase, a suitable initialisation of the tracking phase still presents considerable problems. Part of these difficulties are related to the appropriate choice of assumptions concerning the prevailing illumination of the recorded scene. We present an approach to automatically detect elongated shadow-casting structures in the scene and exploit these structures to automatically distinguish between directed and diffuse illumination. Extended experiments with real-world traffic scenes illustrate the principal practicality of this approach, but simultaneously reveal unexpected difficulties