The Role of Grouping for Road Extraction

An approach to complete road networks extracted from aerial images is presented. Since low-level road extraction results will generally be fragmented and contain false hypotheses, this task involves selecting the correct roads from the extraction results as well as connecting them to construct the road network. Since the presented approach models the road network as a graph, it can handle different kinds of low-level representations, e.g., road segments given by lines or polygons. A flexible scheme to evaluate the quality of the connection hypotheses based on fuzzy set theory is introduced. The problem of connecting the road network is reduced to selecting a few salient nodes in the graph and finding an optimal path between them. Salient nodes are, for instance, the points where a road crosses the image border or points at the ends of very long road segments. Practical results show the validity of the approach.