A structural knowledge-based vehicle recognition method is modified yielding a new probabilistic foundation for the decisions. The method uses a pre-calculated set of hidden line projected views of articulated polyhedral models of the vehicles. Model view structures are set into correspondence with structures composed from edge lines in the image. The correspondence space is searched utilizing a 4D Hough-type accumulator. Probabilistic models of the background and the error in the measurements of the image structures lead to likelihood estimations that are used for the decision. The likelihood is propagated along the structure of the articulated model. The system is tested on a cluttered outdoor scene. To ensure any-time performance the recognition process is implemented in a data-driven production system.