A Novel Up-To-Down Algorithm for Road Extraction

Abstract

On the basis of the symmetric axis transform (SAT), the constrained Delaunay triangulation (CDT) technique, and a split-and-merge approach, a new ‘up-to-down’ algorithm is developed to identify various roads from aerial images. The contours (shapes) of all the potential road regions that are segmented from aerial images are represented by their SAT for decomposing these regions into parts, so that the linear features of the shapes such as the length, width and curvature can be quantitatively measured on the parts, while CDT technique is applied to implement the SAT in discrete domain. From the CDT, the length and width for each part can be computed, and a split-and-merge algorithm is applied for calculating the curvature of each part. Three rules are proposed in terms of the width, length and curvature to identify roads from the candidature road regions. The study shows that the proposed technique is a promising algorithm for identifying roads from aerial images and appears to be superior to the existing methods in that it can extract road networks under complex background with less artifacts. Moreover, as the symmetry based method can quantitatively describe roads, it is also a useful tool for road vectorization.