Quasi-real- time 3D reconstruction from low-altitude aerial images

Generating large-scale virtual 3D models from real-world images is of major interest for many application domains, like urban planning and surveying, and has attracted the computer vision and photogrammetry communities since decades. However, most of the approaches that have been developed are rather restricted since they involve time-consuming manual modeling. Therefore, often the reconstructed scenes are already outdated when they become available for deployment. One recent approach to address this issue is to use Miniature Unmanned Aerial Vehicles (MUAVs) equipped with digital cameras. MUAVs can be used to capture sequences of low-altitude aerial images, which may cover areas under multiple viewing angles with high ground resolution. Using advanced network technology these images are transferred continuously in quasi-real-time, i.e., within a few seconds, to a mission control ground station where they can be processed for immediate display. For this new approach novel photogrammetric processing techniques are needed in order to handle such dynamic and continuously updated low-altitude images. In this paper we present our multiple view 3D reconstruction workflow and introduce a modification to the traditional image registration, which support incremental refinement of the 3D structure and aim at meeting the performance requirements.