%0 Generic
%A Schmude, Naja von
%D 2017
%F heidok:23524
%K Relative Pose
%R 10.11588/heidok.00023524
%T Visual Localization with Lines
%U https://archiv.ub.uni-heidelberg.de/volltextserver/23524/
%X Mobile robots must be able to derive their current location from sensor measurements in order to navigate fully autonomously. Positioning sensors like GPS output a global position but their precision is not sufficient for many applications; and indoors no GPS signal is received at all. Cameras provide information-rich data and are already used in many systems, e.g. for object detection and recognition. Therefore, this thesis investigates the possibility of additionally using cameras for localization. State-of-the-art methods are based on point observations but as man-made environments mostly consist of planar and linear structures which are perceived as lines, the focus in this thesis is on the use of image lines to derive the camera trajectory.    To achieve this goal, multiple view geometry algorithms for line-based pose and structure estimation have to be developed. A prerequisite for these algorithms is that correspondences between line observations in multiple images which originate from the same spatial line are established. This thesis proposes a novel line matching algorithm for matching under small baseline motion which is designed with one-to-many matching in mind to tackle the issue of varying line segmentation. In contrast to other line matching solutions, the algorithm proposed leverages optical flow calculation and hence obviates the need for an expensive descriptor calculation.    A two-view relative pose estimation algorithm is introduced which extracts the spatial line directions using parallel line clustering on the image lines in order to calculate the relative rotation. In lieu of the "Manhattan world" assumption, which is required by state-of-the-art methods, the approach proposed is less restrictive as it needs only lines of different directions; the angle between the directions is not relevant. In addition, the method proposed is in the order of one magnitude faster to compute.    A novel line triangulation method is proposed to derive the scene structure from the images. The method is derived from the spatial transformation of Plücker lines and allows prior knowledge of the spatial line, like the precalculated directions from the parallel line clustering, to be integrated. The problem of degenerate configurations is analyzed, too, and a solution is developed which incorporates the optical flow vectors from the matching step as spatial points into the estimation.    Lastly, all components are combined to a visual odometry pipeline for monocular cameras. The pipeline uses image-to-image motion estimation to calculate the camera trajectory. A scale adjustment based on the trifocal tensor is introduced which ensures the consistent scale of the trajectory. To increase the robustness, a sliding-window bundle adjustment is employed.    All components and the visual odometry pipeline proposed are evaluated and compared to state-of-the-art methods on real world data of indoor and outdoor scenes. The evaluation shows that line-based visual localization is suitable to solve the localization task.