Map-image registration using automatic extraction of features from high resolution satellite images

摘要

In every part of the world the rate of map revision is alarmingly low, when comparedudto the rate of change of many human influenced surface features. Map making isudvery time-consuming and often information used for updates has become historyudbefore the updated map is made available. There is therefore a requirement toudregularly gather up-to-date information about surface features and to incorporateudchanges in maps both quickly and efficiently. Automation of two systems, i.e. theudautomation of map-image registration and then of change detection can fulfill theudrequirements of map revision. This thesis works on the first system.udThe piece of work in this study has looked into a fast and an accurate solution toudregister high resolution satellite images to maps. This will allow changes in groundudfeatures to be used to update maps. Photogrammetric techniques used to updateudmaps have previously shown good results, but they are tedious, time-consuming,udand not beneficial for updating small changes at all.udFeature extraction methods were used in the present study. The system developedudwas designed for automatic extraction of suitable areal features in images.udThe emphasis was on areal features rather than point or linear features becauseudthey have a distinctive shape, and they are extracted easily from vector as welludas raster data. The extraction of suitable polygons, as control information, fromudimages was obtained by using two matching techniques. Patch matching to extractudthe conjugate map and image polygons, and dynamic programming to find theudcorresponding matched boundary pixels of the map and image polygons.udSome matched points were incorrect because of perspective, shadows and occlusions.udA statistical model was developed to remove perspective distortion andudlarge errors. The model demonstrated the removal of erroneous match points, andudselected the good match points and registered the images to maps with a sub-pixeludaccuracy.udA novel aspect of the study is that the automation is achieved with high accuracyudin flat and moderate terrain areas without using height information, as it isudessentially used in photogrammetric techniques.