Selected Approaches for Automatic Road Extraction with Their Models and Strategies

  Some of the early approaches for extracting roads and many later ones aimed at detecting lines from low-resolution imagery. They are exemplified by [8]. Detecting lines can be extended by grouping and network creation [36]. The second common way to extract roads is tracking of roads in high-resolution imagery by extrapolation and matching of profiles after a possibly manual selection of starting points [19]. For a fully automatic extraction, the tracking can be accompanied by an approach to extract roads, which can be used to find starting points [40]. Compared to the above mentioned approaches, the modeling of roads and intersections as well as of objects hindering or supporting its extraction when constructing the network is reasonably improved in [26,24,25]. One step further is taken in [3,29,17]. Here not only the scale-space behavior of roads is utilized, but also the context is divided into spatially more global and more local parts, snakes are used to extract roads and intersections, and the road network is optimized globally. With respect to the complexity of the data input data the more recent approaches, such as [25,3], can handle more complex data.

In contrary to the former approaches, which use only the image as input data, [6] deals with the verification of roads as well as the use of old data from Geographic Information Systems (GIS) to extract new data. For the verification, matching of profiles is used. Inconsistencies in the matching are interpreted in two ways: Either the width has changed or there is a new intersection. The former is checked by matching with other profiles, while for the latter the new roads are tracked. The complexity this approach can deal with lies between medium and high.