Automatic Road Extraction Based on Multi-Scale, Grouping, and Context*

An approach for the automatic extraction of roads ftom digital aeria)-imagery is proposed. It makes use of several versions of the saie aeriaT image with differcnt reso-lutions' Roads ate modeled as a network of intersections and links between these intersections, and are found by a grouping prccess. The context of roads is hierarchically structured into a globd and a Local level. The automatic segmentation of the aerial image into different global contexti, i.e.. rutal, forest' and utban area, is uied to focus the extraction to the most ptomising regions. For the actual extraction of the toads, edges are exttacted in the original high resolution image (0.2 to 0.5 m) and lines are exlracted i-n an image of reduced tesolution. Using,both -resoIution levels and explicit knowledge about toads, hypothese-s for road segments are generated. They are grouped itetatively 'into larger-segments. In addition to the gtoup.ing algorithms, knowle-dge about the \ocal context, e'g., shadows cast by a tree onto a road segment, is used to bridge gaps. To consttuct the road network, finallv intersections are exttacted. Examples and results of an'evaluation based on manually plo-tted reference. daia are given, indicating the potentia) of the approacn. lntroduction In digital photogrammetry, operational automatic solutions existTor eeometiic tasks suchas the measurement of fiducials and matc"hing of homologous points' The latter is used for the reconstruction ofthe relalive orientation, for the generation of digital surface models, or for automatic aerotriangulation' For dala acquisition and update of geographic information systems (crs), the determination of the meaning of individual topographic obiects, e.g., buildings and roads, is necessary' This iemantic tisk still-has to be dbne manually. Because this is time consuming and expensive, automatic solutions are highly welcome. Research on the automatic extraction of topographic obiects lrom aerial and space imagery goes back to the seventies' irlowadays the goal is the update ofcts data. Using existing, albeit ouidated, data can facilitate obiect extraction' However, the approach proposed in this paper is restricted to the extraction of objectt (tte."' roads) without GIS data. There are several reasons for this. First, automatic extraction without a ptiori information shows the potential and deficits of an extraction scheme much better than a cIs-based extraction, because it only relies on the given model and strategy, and therefore can deepen the understanding of the problem. Second, the extraction of new objects is posslble only in this manner and is needed for GIS update in any case. Third, to make the system reliable, it is wise to base the decision about an obiect on new imagery and not on old cIS data. Nevertheless, work on GlS-based extraction of roads (de Gunst and Vosselman, 1997;Bordes et al ',1'gg7; Plietker, 1997) is useful, and has been carried out also within our approach (Wiedemann and Mayer, 1996)' fhe most common techniques for road extraction in images with low resolution are the detection and following of^ linei. tn high resolution, matching of profiles and detection of roadsides, i.e., (anti-)parallel edges, are used' The different approaches apply specialized algorithm_s and additional knbwledge, e.g., geometric constraints. The main criterion to classify the extraction schemes is human interaction. In semiautomitic approaches, an operator provides, for example, starting pointl and starting directions for road_following (McKeown and Denlinger, 19BB; Vosselman and de Knecht, 1995). In Merlet and Zerubia (19s0), points along a road are measured and the algorithm finds the road, i.e., a line which connects these pointi. If more than one image is used, this can also be done in-sl (Griin and Li 1ss7). The advantage of the approaches with multiple points is that the path of the road is more constrained, whiih results in a more reliable handling of critical areas. A similar approach based on so-called "ziplock" snakes is presented in Neu-enschwander ef o1. (1995)' By automatic detection of the seed points, semi-automatic schemes can be extended to (fully) automatic ones. An automatic approach is described in Barzohar et al.(19s7). The selection of itarting points is based on gray-value histograms. Fur-. ther assumptilons about geometry and radiometry are modeled by a Markov stochastic plocess. Road extraction is performed by dynamic programming. In Ruskon6 et 01. (19s4), the centers oi elbngatediegions are detected using a watershed transform of the giadientlmage. Starting from these points, the homogeneity ol the road surface in the images is used to extract road segments. Using geometrical constraints, connection hypothe,rib"t*"".t toa?legments are checked, and a road network is constructed. Similaily to the approach proposed in this article, Trinder and Wang (rdsg) extract roads using different resolutions and grouping. Wherirelaiiois between roads and other objects, e'g., vehicles, buildings, or trees, are neglected, a reliable extraction is "This is a modified version of an article published in German in Photogrammetrie-Fernerkundung-Geoinformation (PFG), No 1/s9' p p . 5 1 7 , 1 9 9 s . A. Baumgartner, H. Mayer, and H. Ebner are with the Lehrstuhl fiir Photigrammetrie und Fernerkundung, Technische Universitdt Miinchen, D-802s0 Munich, Germany ({albert} {helmut}{ebn}@photo.verm.tu-muenchen. de)' C. Steser and W. Eckstein are with the Forschungsgruppe Bildversteiren tFG BV), Informatik IX, Technische Universitdt Miinchen, D-80290 Munich, GermanY ({stegerc} {eckstein}@informatik. tu-muenchen. de)' PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING Photogrammetric Engineering & Remote Sensing Vol . os, No. 7, fu ly 1999, pp. 777-785. oogs-'t11.2 / 99 / 6507 -7 7 7 $3.OO / O O 199s American Society for Photogrammetry and Remote Sensing