Robust grouping of intensity changes in outdoor scenes by a histogram and graph based method

The problem of grouping the gray value variations at the border of complex structured surfaces in outdoor scenes is considered in order to get a simple symbolic representation of such boundaries. Our approach is based on a combination of histogram und minimal spanning tree (a graph-theoretical method) evaluation. In this way the direction and magnitude of the gray value gradient and additionally the spatial distance of contour points are considered in a common framework. After cutting the tree at some edge joining "inconsistent" clusters, which can be selected by combining the mean and standard deviation of all tree edges, the subtrees are symbolically described by their principal axes of inertia. From many successfull applications, the examples presented in this contribution demonstrate that the largest subtree is appropriate in order to localize, e.g., road boundaries.