Road Extraction Based on Line Extraction and Snakes

The extraction of roads from aerial and satellite images is an important task within cartography and planning of new road networks. The automation of this task is highly motivated by the expected increase of the speed and the precision of extraction. The present work considers automatic road extraction from single aerial images of high resolution. It is based on two previously developed approaches: the rst one is the di erential geometric approach for extraction of linear features; the second is the contour extraction based on Active Contour Models, also called snakes. Whereas the rst approach is fully-automatic, the second was previously mostly used for semi-automatic tasks which require the control of a human operator. This work combines both techniques and adapt them in a fully-automatic method for road extraction. According to the used strategy, the hypotheses for most salient roads in images of rural scenes are generated and veri ed rst. Then, based on the ends of extracted roads, the hypotheses for other roads are stated. The developed snakebased technique for the veri cation of these hypotheses enables the recognition of partially occluded and shadowed roads as well as some roads passing through road crossings. The presented results of the developed approach show that the reliable extraction of roads whose images are disturbed by surrounding objects is in many cases possible without the explicit information about these objects. This is a great advantage since the automatic recognition of buildings, vegetation, etc. is a very complicated problem by itself. i V agextraktion baserad p a linjeextraktion och \snakes" Sammanfattning V agextraktion fr an ygoch satellitbilder ar en viktig uppgift inom kartogra och planering av nya vagar. Automatisering av den har uppgiften forvantas medfora okad hastighet och okad noggrannhet vid extraktionen. Det har arbetet behandlar automatisk vagextraktion fr an enstaka ygbilder med hog upplosning. Arbetet baseras p a tv a metoder for bildsegmentering: den forsta ar di erentialgeometrisk linjeextraktion; den andra ar konturextraktion med \snakes". Den forsta metoden ar helt automatisk medan den andra anvants tidigare i samband med manuell styrning. Det har arbetet kombinerar b ada teknikerna och utvecklar en automatisk metod for vagextraktion. I det forsta steget extraherar man de mest framtradande vagarna. Darefter genererar man hypoteser for andra vagar som forbindelser mellan redan upptackta vagar. Den utvecklade snake-baserade tekniken gor det mojligt att veri era hypoteserna och extrahera skuggade och delvis skymda vagar samt vissa vagar som passerar genom vagkorsningar. Resultaten visar att p alitlig vagextraktion av skuggade och delvis skymda vagar i m anga fall ar mojlig utan kunskaper om objekt runt omkring vagarna. Detta ar en stor fordel eftersom automatisk extraktion av objekt som byggnader och vegetation ar i sig sjalv ett mycket komplicerat problem. iii Foreword The present report is a master thesis at the department of Numerical Analysis and Computing Science (NADA) at the Royal Institute of Technology (KTH), Stockholm, Sweden. The work was performed at the Chair for Photogrammetry at the Technical University M unchen (TUM) and was a part of the road extraction project. The author would like to thank his supervisors Prof. Tony Lindeberg (KTH) and Dr. Helmut Mayer (TUM) for their support of the author's wish to investigate the fascinating eld of computer vision. The author is especially grateful to Dr. Helmut Mayer for his invaluable help in numerous aspects during the whole work. The author also thanks Dipl.-Ing. Albert Baumgartner, Dipl.-Ing. Carsten Steger, Dipl.-Ing. Christian Wiedemann, Dr. Wolfgang Ecksteinas and all the members of the Chair for Photogrammety, TUM, who were always ready to help and advise. v