Journal of Imaging Science and Technology ❝ R 56(3): 030504-1–030504-11, 2012. c  Society for Imaging Science and Technology 2012 Urban Area Road Boundary Extraction from a True Orthoimage using Ribbon Snakes Junhee Youn Korea Institute of Construction Technology, 283 Goyangdae-Ro, Ilsan-Gu, Goyang-Si, Gyeonggi-Do, Korea Gihong Kim Department of Civil Engineering, College of Engineering, Gangneung-Wonju National University, 120 Gangneung Daehangno, Gangneung, Korea E-mail: ghkim@gwnu.ac.kr Abstract. This article describes automatic road boundary extraction from a high resolution true orthoimage using the Ribbon Snakes algorithm. We assume the existence of prior information for the rough range of road widths in the scenes and road centerline data. Previous works on road boundary extraction have been focused on rural areas, using the same approach. Applying the Ribbon Snakes algorithm to an urban area, we encounter multiple-local-minima problem due to cars and lane markings. We overcome this problem by repeating the optimization of the Ribbon Snakes model at different widths. With the existing road centerlines, fixed width Ribbon Snakes (FXWRS) is applied and its total energy after optimization is stored. After changing the road width, FXWRS is reapplied and the total energy is again stored. By comparing the total energies for each of the road widths, we determine the optimum width that produces the least total energy. Applying FXWRS with the optimum width and road centerline, refined road boundaries are obtained. We show the feasibility of this approach by comparing results with ground truth. ❞ c 2012 Society for Imaging Science and Technology. [DOI: 10.2352/J.ImagingSci.Technol.2012.56.3.030504] INTRODUCTION Road boundary extraction from imagery has become an important research topic in recent decades, with the increas- ing demand for accurate spatial information for geographic information systems (GIS). Many GIS data sets only include road centerline information, and this is appropriate when the map usage is large scale (e.g., calculation of the time for driving to a destination). But for small scale usage of GIS, such as in visually supporting drivers and autonomous driving, double-line roads, which present the road bound- aries, are essential. Also, highly accurate road boundaries and centerlines are often needed in vehicle detection and traffic analysis. 1 The usual strategy for road boundary (or edge) extrac- tion from high resolution satellite or aerial imagery is refining the road edges on the basis of the initial approximations of road centerlines taken from existing road network data or from prior detections by independent means. Heipke et al. detected the road centerlines in the low resolution images with global and local thresholding, and refined the road Received Jul. 29, 2011; accepted for publication Jun. 15, 2012; published online Oct. 22, 2012. 1062-3701/2012/56(3)/030504/11/$20.00 boundaries from high resolution images, grouping parallel edges around the centerlines. 2 Baumgartner et al. and Mayer et al. expanded this multi-resolution approach 3 , 4 combining it with Ribbon Snakes 5 optimized by using Ziplock Snakes. 6 Laptev et al. followed a multi-resolution and Ribbon Snakes strategy to detect road centerlines and boundaries, bridging the road segments in shadows or partially occluded areas. 7 In suburban (residential) areas, initial approximations of the road centerlines can be obtained by connecting road intersections. Koutaki et al. detected road intersections using an intersection model matching algorithm from the road-classified binary image. 8 With initial approximations connecting detected road intersections, the authors extract road boundaries by applying Ribbon Snakes, whose image energy term is defined as the difference between the inner and outer regions of the ribbon. The above approaches have concentrated on rural areas or suburban areas, and few research groups work on road boundary extraction in an urban environment. This is because of the complexity of urban areas. Road boundaries were refined, by combining a multi-resolution approach and wavelet transforms, from the satellite image by Couloigner and Ranchin. 9 In this article, the initial information about the road consists of four corners of a rectangular road segment at the different spatial resolutions. On the basis of the initial approximations, road boundaries are extracted using approximations at different spatial resolutions, and road strips are detected by using wavelet coefficient images at characteristic scales and the extracted boundaries. As stated before, the Ribbon Snakes algorithm was used for road boundary extraction. However, it was developed mainly for applications in rural areas due to the local minima problem. If the initial approximations of Ribbon Snakes are not located close to the contour of interest, it may trapped by other local minima caused by lane markings or cars. This is the motivation of our work. Péteri and Ranchin have shown that such problems can be overcome by using a multi-resolution approach and ‘‘Double Snakes’’, which is composed of two lines evolving jointly to maintain parallelism. 10 In this article, initial approximations of the road location in coarse imagery were derived from the road segments in existing road network data. The optimized Snakes solution for the J. Imaging Sci. Technol. 030504-1 May.-Jun. 2012