Visual location recognition in man-made and natural environments
Georges Baatz · Repository for Publications and Research Data (ETH Zurich) · 2013
This thesis addresses the problem of estimating the position where a given photograph has been taken.The first system targets images taken in inhabited areas.Building facades, which cover a significant proportion of the field of view, are used to recognize the building, and thus the location.The presented approach builds on panoramic images captured by a surveying vehicle systematically driving through all streets and coarse geometric models of the buildings.Both database and query images are warped into canonical views to reduce the variations that images from arbitrary viewpoints naturally exhibit: The photos are warped into orthographic views and rotated such that the real-world vertical coincides with the picture's vertical.For the survey imagery this can readily be done using the building models, and for the queries it is achieved with the help of vanishing points.Undistorting the images simplifies similarity calculations significantly, and allows for further improvements such as more discriminative features and a geometric verification stage that is both stricter and faster.The system is validated on a city-scale database.Experiments show that the proposed approach reaches a performance comparable to earlier systems, that could only handle much smaller datasets.A second system deals with photographs of landscapes.The skyline, i.e. the apparent boundary between sky and ground is characteristic to the location and is the basis for localization.The reference data consists of a digital elevation map, which can predict