Comparing stereo 3-d with view-based vehicle recognition

Two approaches to polyhedron model-based vehi-cle recognition are compared. Both are structural techniques and work on image sequences. The first approach uses stereo triangulation of simple L-shaped structures and matches parts of the 3-d model with the 3-d structures. It assembles the final recognition by following the part-of hierar-chy and the geometric constraints of the target object in the 3-d scene. In contrast to this the second approach constructs a set of 2-d view-models from the 3-d model by off-line hidden-line projection. These are compared with L-shaped structures of the image by means of a generalized Hough-accumulator. Both approaches are tested on a common sequence of oblique views of a truck. These give wide-angle perspectives, which actually do not suit the view-based approach. With respect to this abuse the technique still gives a satisfactory result, advocating its robustness. The result of the 3-d technique is much more neatly aligned.
Michaelsen E, Wankmueller U, Stilla U (2002) Comparing stereo 3-d with view-based vehicle recognition. In: Greiner G, Niemann H, Ertl T, Girod B, Seidel HP (eds) Vision, modeling, and visualization, VMV2002. Berlin:Infix, 347-354
[ ]