Automatic building detection from aerial images for mobile robot mapping

To improve mobile robot outdoor mapping, information about the shape and location of buildings is of interest. This paper describes a system for automatic detection of buildings in aerial images taken from a nadir view. The system builds two types of independent hypotheses based on the image contents. A segmentation process implemented as an ensemble of SOMs (Self Organizing Maps) is trained and used to create a segmented image snowing different types of roofs, vegetation and sea. A second type of hypotheses is based on an edge image produced from the aerial photo. A line extraction process uses the edge image as input and extracts lines from it. From these edges, corners and rectangles that represent buildings are constructed. A classification process uses the information from both hypotheses to determine whether the rectangles are buildings, unsure buildings or unknown objects.