Modeling contextual knowledge for controlling road extraction in urban areas

This paper deals with the role of context for automatic extraction of man-made structures from aerial images taken over urban areas. Due to the intrinsic high complexity of urban scenes we propose to guide the extraction by contextual knowledge about the objects. We represent this knowledge explicitly by a context model. Based upon this model we are able to split the complex task of object extraction in urban areas into smaller sub-problems. The novelty presented in this contribution mainly relates to the fact that essential contextual information is gathered at the beginning of the extraction, thus, it is available during the whole extraction, and furthermore, it allows for automatically controlling the extraction process: for data consistency reasons, we use the imagery as the only source for both gaining contextual information and extracting roads. Advantages and remaining deficiencies of the proposed strategy are discussed.