Page - 132 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 5. Processing pipeline of the point cloud fusion: (1)Raw data from dense image matching (50.64 M points), (2) fused point cloud (1.73 M points), (3) discarding weights smaller thanα=30 (0.47 M points), (4)mesh generation, and (right side)merged surface tiles. Regarding the oblique dataset, best results can be achieved by neglecting points with TV class 1. By doing so, execution time is speedupbyafactorof2.2.Compared to the rawpoint cloud the fusionprocedure reducesnoisewhile improving the accuracy of the point cloud (see Fig. 6). A visual assessment shows that the fused point cloud including all TV classes and applying weights produces the best results regarding completeness and outliers (see Fig. 7). As expected, roof Fig. 6. Comparison of the main school facade before and after fusion procedure (cf.Fig.5):MeandeviationbetweenDSMderived fromterrestrial laser scanner data and point cloud (top), and standard deviation of the point clouds DSM representing the level of noise (bottom). Fig. 7. Taking all TV classes into account produces point clouds containing less outliers (left), in contrast to point clouds restricted to TV classes> 1 (right). structures and other nadir oriented faces are reconstructed with the highest precision. Table II shows that in all cases the precision of the point cloud can be improved while decreasing redundant information. TABLE II COMPARISON OF TEST AREAS BEFORE AND AFTER THE POINT CLOUD FUSION. Density RMSE Fused Mean Fused Std. Dev. [pnts/m2] PC-TLS [m] PC-TLS [m] of DSM [m] Tower South (raw) 2345.9 0.378 0.051 0.538 Tower South (fused) 49.4 0.204 0.003 0.087 Tower North (raw) 1781.4 0.427 -0.222 0.447 Tower North (fused) 45.3 0.195 -0.052 0.071 Tower West (raw) 3570.8 0.350 0.237 0.499 Tower West (fused) 62.7 0.256 0.152 0.155 Roof (raw) 13864.2 0.150 -0.023 0.218 Roof (fused) 178.7 0.122 0.028 0.105 B. Nadir Aerial Imagery The nadir image dataset covers an area of approximate 1.5 × 1.7 km2 in the city of Munich. The dataset was acquired by a DMC II 230 megapixel aerial image camera with a spatial resolution of 10 cm and consists of 15 panchromatic images. As depicted in Fig. 8, facade information can be reconstructed by utilizing the proposed fusion routine. Due to the wide angle of the aerial camera, enough information is captured to produce 3D city models from nadir aerial imagery. V. CONCLUSION A novel method for fusing 3D point clouds was presented. The underlying point clouds originate from stereo matching of aerial images and were enriched by the calculation of surface normals and a classification of the disparity maps into quality classes. The proposed filtering method then fused the point cloud in direction of the surface normals and used a weighting based on the classification. Evaluation to ground truth data showed the increased quality of the fused point cloud while reducing the redundancy. Overall, 132