Advisors: Norbert Pfeifer, Camillo Ressl
Historical images offer us a unique insight into the past. As such, they have become important sources for documenting landscape changes. Among historical images, oblique images are less used by researchers than other sources, such as aerial nadir or satellite images. Since they were acquired completely unstructured by amateurs with unknown cameras and are scattered over multiple archives without common metadata, integrating them into spatial analysis is challenging and time-consuming. To overcome those challenges and hence, exploit the full potential of these images, photogrammetry can play a vital role. 
Especially, the unknown interior and exterior orientation is preventing further exploitation in two ways: First, images in archives cannot be searched spatially. Second, the precise estimation of the interior and exterior orientation through spatial resection, a prerequisite for documenting landscape changes from single oblique images through monoplotting, requires ground control points whose identification is time-consuming. Therefore, accurately estimating the unknown camera parameters with minimal manual input is one of the key challenges for opening up this immense resource for research. 
With the known camera parameters, monoplotting becomes possible. However, with none of the existing monoplotting solutions, the uncertainty of the monoplotted object points can be estimated. While the overdetermined spatial resection also estimates the uncertainty of the camera parameters, it is still unclear how these uncertainties propagate to the monoplotted object points. Only with this additional information, thorough integration of monoplotted features into spatial analysis be¬comes possible. Otherwise, it remains unclear whether the accuracy of monoplotted points suffices to quantify observed landscape changes. 
Hence, within this work, we specifically addressed those two challenges, which we believe are the main bottlenecks for exploiting this unique resource. To efficiently estimate the unknown camera parameters for whole image collections, we developed an automatic image orientation approach based on the visible horizon. The horizon, in contrast to the remaining parts of the images, has undergone less significant changes over the last century and hence, can be considered a stable feature for matching. However, this approach requires extensive preprocessing, which might not be suitable for orienting just a few images. Therefore, we further developed a semi-automatic image orientation approach using deep learning based feature point descriptors to find corresponding points between historical oblique images and rendered 3D scenes. To advance the subsequent monoplotting, we developed an approach to estimate the uncertainty of monoplotted object points and investigated cases (i.e., silhouettes), where the uncertainty estimation does not yield correct esti¬mates. Lastly, we integrated the semi-automatic image orientation and uncertainty estimation into our monoplotting tool moniQue, which was directly integrated into QGIS to bring the processing of oblique images as close as possible to the potential users. 
Through all these achievements, we believe that this work will help to bridge the gap for researchers to integrate historical oblique images into their research. And with that, we might help to close the observation gap currently present in research related to landscape changes.