Automatic Urban Road Extraction from Digital Surface Model and Aerial Imagery

Automatic linear feature extraction has a long history but still it is one of the challenging topics in photogrammetry and computer vision. In this paper, we describe an automatic road extraction process suitable for urban areas with high buildings. The proposed road extraction process has roughly three steps. The first step is generating an obstacle map to get road primitives from the digital surface model (DSM). In this step, it is assumed that roads are on the ground surface. The second step is road centerline extraction from the road primitives. Based on the assumption that roads are straight in dense urban areas, the Hough transformation is used to detect road segments and the line parameters are adjusted by the least squares method. In the third step, based on the extracted road centerlines, road edges are extracted by an adaptive snake algorithm from a true orthophoto. In this procedure, seed points for the adaptive snake algorithm are automatically generated by using the result of the previous step. The proposed road extraction process is tested on the downtown area of San Francisco, California.