Road Network Extraction using Edge Detection and Spatial Voting Beril Sırmac¸ek and Cem ¨ Unsalan Computer Vision Research Laboratory Department of Electrical and Electronics Engineering Yeditepe University ˙ Istanbul, 34755 TURKEY Beril.Sirmacek@dlr.de, unsalan@yeditepe.edu.tr Abstract Road network detection from very high resolution satellite images is important for two main reasons. First, the detection result can be used in automated map making. Second, the detected network can be used in trajectory planning for unmanned aerial vehicles. Al- though an expert can label road pixels in a given satel- lite image, this operation is prone to errors. Therefore, an automated system is needed to detect the road net- work in a given satellite image in a robust manner. In this study, we propose a novel approach to detect the road network from a given panchromatic Ikonos satel- lite image. Our method has five main steps. First, we apply a nonlinear bilateral filtering to smooth the given image. Then, we extract Canny edges and the gradient information as local features. Using these local fea- tures, we generate a spatial voting matrix. This voting matrix indicates the possible locations of the road net- work pixels. By processing this voting matrix in an it- erative manner, we detect initial road pixels. Finally, we apply a tracking algorithm on the voting matrix to detect the missing road pixels. We tested our method on various satellite images and provided the extracted road networks in the experiments section. 1 Introduction Very high resolution satellite images, like Ikonos and Quickbird, can be used to generate land maps. The first part of land map can be taken as the road network. The resolution of the Ikonos and Quickbird satellite images are suitable for applying computer vision algorithms to extract the road network. Unfortunately, well-known computer vision algorithms are not suitable alone for automatically extracting the road network from a given image due to several reasons. Road segments can be oc- cluded by other nearby objects like buildings and trees. They may also have different colors. Their widths may change. Moreover, junctions of unknown number of roads and roundabouts may increase the difficulty of the problem. Therefore, advanced methods are needed to extract the road network from very high resolution satellite images. In the literature there are several methods to detect the road network from a given satellite or aerial im- age. Yang and Wang [10] developed a method to detect main roads from satellite images. First, they detected road primitives such as straight lines and homogenous regions. Then, they linked detected primitives. Unfor- tunately, their method can not detect urban roads and occluded road segments. Ma et al. [5] detected par- allel edges in panchromatic ETM (Enhanced Thematic Mapper) images to locate road segments. They linked discontinuous road segments using perceptual organiza- tion rules. Rianto and Kondo [6] proposed an approach to detect main roads in SPOT satellite images. For this purpose, they detected Canny edges and classified straight line segments using the Hough transform. They assumed straight and parallel line segments as roads. Unfortunately, this approach can not be sufficient alone to detect curvilinear and complex roads in urban scenes. There are also semi automatic techniques to detect road segments [1, 8]. Baumgartner et al. [2] and ¨ Unsalan and Boyer [9] provide excellent surveys on road detection in aerial and satellite images. In this study, we propose a novel method to detect the road network in very high resolution panchromatic Ikonos satellite images. Our method is based on five main steps. First, we apply nonlinear bilateral filtering to smooth the given image [4]. Then, we extract Canny edges and the gradient information as local features [3]. Using the extracted local features, we generate a spatial voting matrix. This voting matrix indicates the possi- 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.762 3105 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.762 3117 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.762 3113 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.762 3113 2010 International Conference on Pattern Recognition 1051-4651/10 $26.00 © 2010 IEEE DOI 10.1109/ICPR.2010.762 3113