Urban terrain reconstruction has many applications in areas of civil engineering, urban planning, surveillance and defense research. Therefore the needs of covering ad-hoc demand and performing a close-range urban terrain reconstruction with miniaturized and relatively inexpensive sensor platforms are constantly growing. Using (miniaturized) unmanned aerial vehicles, (M)UAVs, represents one of the most attractive alternatives to conventional large-scale aerial imagery. We cover in this paper a four-step procedure of obtaining georeferenced 3D urban models from video sequences. The four steps of the procedure – orientation, dense reconstruction, urban terrain modeling and geo-referencing – are robust, straight-forward, and nearly fully-automatic. The two last steps – namely, urban terrain modeling from almost-nadir videos and co-registration of models 6ndash; represent the main contribution of this work and will therefore be covered with more detail. The essential substeps of the third step include digital terrain model (DTM) extraction, segregation of buildings from vegetation, as well as instantiation of building and tree models. The last step is subdivided into quasi- intrasensorial registration of Euclidean reconstructions and intersensorial registration with a geo-referenced orthophoto. Finally, we present reconstruction results from a real data-set and outline ideas for future work.

Convolutional neural networks, or CNNs, raised the bar for most computer vision problems and have an increasing impact in remote sensing. However, since they usually contain multiple pooling layers, detection of exact borders of small objects at their original resolution remains yet a challenging topic. Additionally, efforts are being made to reduce the amount of training data. In this paper, we investigate the potential of fully convolutional neural networks (FCNs) for individual vehicle detection in combined elevation and optical data using relatively few training samples. By the proposed multibranch CNN, we combine object recognition within a deep learning framework with the object segmentation at a high resolution, for which two CNN branches are employed. Data fusion is accomplished with a pseudo-Siamese approach. The pixelwise classification likelihood, also referred to as heatmap, is harmoniously postprocessed by a vectorization module, which is based on the minimum bounding rectangle (MBR) extraction and allows for delineation of groups of vehicles. Two methods were developed in which MBRs are supported either by pairs of parallel lines or by region growing. Our approach allows efficient training with few training samples, while delivering high-quality detection results and good computational performance. In our detailed evaluation, we investigate the benefits of data fusion and compare our approach to other state-of-the-art networks. Different datasets were used, containing optical images and elevation data, derived either from airborne laser scanning or from photogrammetric reconstruction. The obtained results are very promising with F$_1$ scores up to 97%.