This thesis proposes a new approach for the automatic extraction of roads from digital aerial imagery. It focuses on fully automatic extraction and utilizes an explicit object model. Compared to other approaches the most prominent features of this thesis are the exploitation of the scale-space behavior of roads and the use of contextual information by means of global context regions and local relations between roads and other objects. The approach aims at road extraction in open rural areas. Panchromatic aerial images with a pixel size of approximately 0.2 to 0.5 meter on the ground serve as input data for the automatic extraction. The proposed approach makes use of several versions of the aerial image with different resolution. Roads are modeled as a network of intersections and links between these intersections. For different so-called global contexts, i.e., rural, forest, and urban area, the model defines relations between background objects and road objects. These relations, e.g., that a tree casts a shadow on a road-segment, determine so-called local contexts. These local contexts are modeled differently depending on the global context regions. An automatic segmentation of the aerial image into different global contexts by means of texture classification is used to focus the extraction on the most promising regions. Additionally, it allows to predict in which parts of the image the results will be most reliable. For the actual extraction of the roads edges are extracted in the original high resolution image (pixel size 0.2-0.5 m) and lines in an image of reduced resolution (pixel size 2-4 m). Using both resolution levels and explicit knowledge about roads hypotheses for road-segments are generated. They are grouped iteratively into longer segments. In addition to pure grouping criteria also knowledge about the local context and so-called "Ribbon-Snakes" are used to bridge gaps. For the construction of the road network intersections are extracted. The examples presented and the results of an evaluation based on manually plotted reference data show the efficiency of the approach.
«