Strategy

  The parts of the strategy are organized in the same way as in Subsection 3.1.

The general parts of the strategy are:

• Appearance-based approaches [26] avoid the explicit transition from image model to object model [22]. With them, objects like trees can be extracted, for which the modeling is quite difficult by other means because they have a varying appearance. They also can be used for the extraction of cars on roads or details on the rooftop when the resolution is close to the point where these objects cannot be extracted any more.
• Grouping [28,30], i.e., the search for geometric/topologic regularities, allows an algorithm to focus on parts of objects and therefore limits the search space. An often encountered problem is that the regularities specified are not strict enough to ensure a reliable extraction. Grouping should therefore be accompanied by verification.
• By means of the focus on different scales [3,35,34] the extraction is at the same time sped up and improved. By using multiple scales, one can start with reliable structures in coarse scale and use them to focus the extraction on the specific areas and object types in fine scale. In many cases, instead of changing the scale in the image by means of scale-spaces [15], image pyramids can be used, which significantly accelerates the processing of the coarse scale.
• Hypotheses generation and search/resegmentation based on spatial context [9,16] is done by predicting an object given another object with a spatial relation to the first object. Many objects receive their semantics only in this way, which is especially true if they cannot be recognized, or are at least hard to extract by themselves.
• When focusing on contexts [3] the distinction into global and local context is used for a further improvement of hypotheses generation and search/resegmentation. In many cases it is useful to first segment the image into the global contexts and only then start the extraction of the objects in the easiest or most promising global context. For roads, these are, for instance, the open_rural areas, in which the objects and the local contexts made up from them are analyzed. Objects in the local context, such as trees or shadows, can prevent the extraction of roads. Other objects, such as cars, can help to validate roads.
• The Generation of evidence from structures of parts/substructures [26,3,9,23] improves the probability of hypotheses. Here, it is assumed that substructures cannot be extracted directly in many cases. However, if there is a hypothesis about the object to be extracted, its spatial constraint makes the extraction of the substructures possible. For roads, single objects, such as markings, cannot be interpreted by themselves (e.g., faint bright lines), but their arrangement make their semantics and, at the same time, the semantics of the object itself clear.
• Balancing image information versus geometric model in an automatic process [11,17] enables a geometric improvement of objects with already clear semantics, but weakly defined outlines. Typical examples are snakes [13]. Recent results on the extraction of roads in shadowed regions [17] show that snakes are also useful to extract objects when only a stabilized geometry makes the extraction of useful image features possible.
• The fusion of data and algorithms [9] comprises for the data not only color but also multi-spectral images and images from different sensors. Although, the color in images is not stable due to the indirect lighting of shadowed objects, color images are for instance useful for limiting the search space of road extraction by using the fact that many roofs are red and therefore cannot be roads. The fusion of algorithms is a very general technique which can be used for different areas. Examples are the treatment of scale transitions or of different kinds of viewpoint of the image function, e.g., various types of definitions of edges or regions.
• The utilization of GIS [6,23,4] allows a significant speed-up of the interpretation by focusing the extraction onto relevant areas. This avoids the large effort for a complex generation of hypotheses from the images.

Specific parts of the strategy for roads are:

• Road tracking [6,1,19,31] is based on profiles or a Markov-Random-Field (MRF) based road model combined with a Kalman Filter. Therefore, it makes full use of the information about location and direction of the road, but its drawback is its high computational effort.
• By grouping for the construction of the road network [36,7] a local formation of intersections and a closing of gaps is achieved.
• The grouping based on the network [29,25], which depends on the global structure of the network, broadens the previous item by introducing paths which are optimum in terms of the whole network.