Reconstruction of geometric road data using remotely sensed imagery

Abstract: The objective of this thesis is to find an automatic methodfor accurate and robust reconstruction of geometric road data.Remotely sensed images serve as the main source during thereconstruction process. Rapidly increasing requests foraccurate and actual road data, mainly driven by commercialinterests such as route optimisation, car navigation, andtelecommunication topics, stresses the need for automation ofmapping procedures. As shown in this thesis, two majoradvantages are achieved if the existence of a geographicaldatabase is taken into consideration during the imageinterpretation procedure. First, the geographical updatingprocedure is simplified and automated while using the locationof old already mapped objects as start positions during theroad database reconstruction. Second, semantic information alsoavailable in the database can be used in the image analysis.For example, road width is a useful parameter forobject-related image operations.The developed road data reconstruction method is dividedinto four subtopics. First, the ortho-corrected image isprocessed using derivatives of a Gaussian filter to enhancelinear structures. The road width, either extracted from thedatabase or estimated by an operator, steers the filteringprocess. Second, road intersections and dead ends, in thisthesis named node points, are determined. Different matchingalgorithms, all based on existing road data and thepre-processed image, are developed and tested. Third, awell-established method, snakes, is used for determination ofthe road delineation between two node points. Drawbacks such ashigh requests for accurate start values and sensitive parametersettings initiated the idea of a road data reconstructionmethod based on Least Squares image matching. The methoddeveloped and tested in this thesis matches the modelled road,representing the road cross-section, and a remotely sensedimage. Two fundamental similarities between the adopted nodepoint and delineation methods are the formulation of anartificial template image describing the road cross-section andthe required image pre-processing. Smoothing constraints areincluded in the iterative delineation procedure to ensure astable solution. Finally, automation of geographical dataupdating requires accuracy estimates of all data sourcesincluded in the process. The method presented in this thesisallows for calculation of statistical measures, where eachvertex of the road segment is quality estimated, useable asinput in a decision procedure for updating.Accuracy evaluations are performed on a test data setincluding results from a manual interpretation and digitisationof SPOT multispectral images. Differential GPS measurementsfrom 27 kilometres of a forestry road network serve asreference data during the comparison between manualinterpretation and the results from the developed roadreconstruction method. The test results from the automated roaddata reconstruction show an improved accuracy compared tomanual interpretation. An analysis of the error componentsindicates a major influence from the image location, 7.5metres, while the contribution from the road reconstructionalgorithm is approximate 4 metres, in this test correspondingto 0.2 pixels.Keywords:object-related image processing, node pointdetermination, road delineation, image matching, accuracyestimates, accuracy study.