Model

  The model is organized into general parts, which are also useful for other objects, especially buildings, and specific parts for roads.

The general parts of the model are:

• Characteristic properties are often the consequence of the function of objects [3,6]. Typical examples for knowledge sources, apart from constraints concerning the usefulness for humans [27], are the construction instructions for different types of roads. For large parts of the knowledge about function, it seems to be enough to take those parts into consideration for modeling, i.e., it is not necessary to integrate them into the system.
• The modeling of material properties [33] makes that the interpretation is not too much affected by sensor characteristics. This is especially important when different kinds of sensors like optical and radar are utilized.
• 2D geometric/topologic regularities [30,28] are used for road pavements as well as for buildings (parallel edges).
• A detailed image model [12,28] with rich attributes and a feature adjacency graph exploits the information contained in the image much better. Especially for matching in more than one image, the probability can be raised significantly compared to approaches based only on one feature type (mostly edges). For roads, the combination of parallel edges and the homogeneity of the enclosed area gives good evidence in many cases.
• Levels of abstraction and scale [3,34] are especially useful for roads and vegetation. For roads in coarse scale, many disturbances are eliminated [18]. The elimination of disturbances helps to bridge gaps and to get a more complete road network. Coarse scales can be generated artificially using scale-spaces. The local modeling of single objects like trees by using appearance-based approaches [22] is exploited in fine scale. This avoids the transition from image space to object space.
• The geometric/topologic neighborhood [9,39,33], i.e., the spatial context, describes the spatial arrangement of objects. For example, intersections or cars have a direct relation to a road, whereas trees or buildings are needed because they cast shadows or occlude roads on one hand, and because they form rows parallel to the road on other hand.
• Global and local context [3,4] subdivides the geometric/topologic neighborhood by a spatial partitioning. Local structures like a tree casting a shadow or a building occluding a road are distinguished from global structures like suburb_urban, forest, or open_rural. The latter ones restrict the frequency and the characteristics of the former ones: For instance in open_rural areas there are few trees casting shadows.
• Structures of parts [24,3,9,23], also called substructures can be used as local evidence for objects. Typical examples include cars on the road, doors, or windows in the wall, as well as dormer windows or vents on the roof. These objects show a characteristic arrangement among each other and regarding the object they are part of. Often it is useful to rectify the images before the extraction of the part structures to get a standardized situation.
• Statistic modeling [14] extends the widely used, more or less functional and deterministic modeling. With probabilistic methods the uncertainty of the data and of the model can be propagated and used for controling the analysis.

The specific parts of the model for road extraction consist of:

• Lines [3,5,20,10,34] are used to model roads in low and medium resolution. For wide roads and for not too high requirements with respect too completeness, e.g., for small scale mapping, this modeling can be enough. Generally, lines, for the delineation of which various criteria exist, are not sufficient to characterize a road in the image.
• The Road pavement [3,25,1,6], which is mostly described in terms of parallel edges (apars) or profiles, is suitable for fine scale. Many roads in the image are adequately modeled this way. Compared to lines, disturbances from other objects and shadows/occlusions have a more significant influence on the extraction.
• The Road network [25,29,36,7,34] consisting of connections and intersections, ties the roads together. By a local bridging of gaps, the network can be improved significantly in many cases.
• By using optimal paths the global network [25,29] extends the local network construction with a criterion of optimality. Not only have the connections to be established but they also have to be optimum in terms of the whole network. For instance, outside of urban areas two neighboring points should not be connected by a path of similar quality. What is not guaranteed by optimal paths is global topological consistency, i.e., it is not checked if roads are dead ends.