Locating buildings in urban area from geometrical information

Abstract

An identification of geographic objects such as buildings from images is a challenging task in image processing, since it can be applied to various applications, e.g. city planning and disaster management. Consequently, an automatic building detection based on line scanning is proposed in this dissertation. For top view image, all possible candidate areas are initially identified on the tested image. To avoid redundant time consuming, the image is chopped into sub-images according to the candidate areas. An appropriate size of the sub-images is approximately estimated by a histogram of candidate area sizes. Then, degrees of angles relevant to the candidate areas are investigated by Hough transform. All significant lines related to the sub-images are subsequently extracted to be used as initial lines. Finally, the rectangle shaped objects are detected by the proposed line scanning algorithm. The experimental results shows that the proposed algorithm can acquire higher performance than Karsli’s method, and can achieve at least 90% of accuracy for object based building detection. For perspective view, the proposed algorithm shows that it can be consistently applied to building detection on the tested image, since it can extract lines that are building components.