Page - 130 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

facade patches where reference data is available. Analysing the DSM derived from the classified pointcloud and the reference data enables us to compute the weights in form of a weighting function. The weighting function is derived by calulating the standard deviation of the flatness error and fitting an exponential function in a least squares manner. The flatness error is defined as the point cloud deviations to a best fitting plane and is also an indicator for the noise of the 3D geometry [1]. Later on, we evaluate the fused pointcloud in a similiar way, to gain insight on the potential and quality of the entire fusion method. Specific information regarding the evaluation routine, selected test areas and datasets are given in Section IV. C. Weighted-Median Based Fusion The concept of median-based fusion originates from fusion algorithms for the generation of 2.5D DSMs. Rothermel et al. [15] adapted the idea by fusing point clouds in 3D space along a defined filtering direction. While for close range datasets the line of sight is suitable as filtering direction, point-wise normals are used for the fusion of aerial datasets. We adapt this fusion strategy using a weighted-median based approach. In a first step, an initial pointsetP is created from the input point cloud by storing the input point cloud in an octree data structure. The pointset P is derived by subsampling the point cloud with the centroid of the points located in a leaf node. In our work the entire fusion process was realized with the aid of the Point Cloud Library (PCL ver. 1.8.0) [16] which also provides a custom tailored octree implementation. As a result of the disparity quality assessment every point possesses a weight representing the quality of the point. We add up the weights of all points located in the same leaf node. Thus, the weight of the initial point p∈P is an indicator for the density and quality of the reconstructed scene. Subsequently, the point cloud is fused using nearest neighbor queries optimized for cylindrical neighborhoods. For every point in the initial pointset P a set of candidate points Q, located in a cylinder with its central axis given by the initial point and its normal, is derived. Points with surface normals diverging more than 60◦ are discarded for further processing. After the candidate pointset Q is detected, the point p is filtered by projecting all candidate points onto the surface normal of the initial point p. Taking the weighted-median of all deviations to the point p yields the new point coordinates. Especially for noisy data further iterations can be inevitable to generate a consistent surface representation. Between every iteration, duplicate points are united to avoid redundant computations. A detailed description of the original fusion routine including the parameters and employed neighborhood queries is given in [15]. In a first iteration, Rothermel etal. [15] includes all points of the input point clouds for the identification of the candidate pointset Q. To speed up further iterations the filtering is restricted to the initial pointsetp∈P solely. In our case, we restrict the filtering of the point cloud to the initial pointset P from the beginning on. We compensate the loss of detail of the input point cloud by approximating the density of the captured 3D scene with the accumulated weight. The final surface representation is derived by discarding points with weights smaller than a defined threshold α. The influence of the threshold is analyzed in Section IV-A. In this way large and highly redundant 3D point clouds can be fused in moderate time (e.g. processing 2.5 billion points on a computer with 16 cores within a single day, resulting in a fused point cloud whose density fits the spatial resolution of the input imagery). IV. RESULTS In this section we discuss results obtained with the pro- posed fusion pipeline. The datasets used for the evaluation are provided by the ISPRS/EuroSDR project on “Benchmark on High Density Aerial Image Matching”2 and consist of one nadir and one oblique dataset. A. Oblique Aerial Imagery The oblique imagery dataset was acquired over the city of Zu¨rich with a Leica RCD30 Oblique Penta camera consisting of one nadir and four oblique 80 megapixel camera heads. While the nadir camera head is pointing downwards, directly towards the earth, the four oblique camera heads are tilted at an angle of 35 degrees, each pointing in a different cardinal direction. The entire datasets comprises 135 images, captured from 15 unique camera positions. While the nadir imagery leads to a Ground Sample Distance (GSD) (i.e. the spatial resolution) of 6 cm the GSD of the oblique views vary between 6 and 13 cm. Reference data captured with terrestrial laser scans provide accurate and reliable information for the evaluation of the datasets. The evaluation was carried out by computing DSM’s of different facade patches distributed over the test area. More information on the image acquisition, benchmark and reference data can be found in [2]. Photogrammetric Processing and Pre-processing. In a first step, the image registration was carried out using the interior and exterior orientation parameters provided along with the image data. Subsequently images are matched in flight direc- tionwithanoverlapof70%, resulting ina totalof314stereo- pairs, containing approximately 10.6 billion points. After the generation of disparity maps TV classes and normal maps are computed. As mentioned in Section IV the weighting function assigns a weight to every TV class which is then used in the fusion process. The derived weighting function is depicted in Fig. 3 and shows that a correlation between TV classes and the geometric precision (i.e. level of noise) can be verified. 2http://www.ifp.uni-stuttgart.de/ISPRS-EuroSDR/ ImageMatching/ 130