Page - 131 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 2. Raw dense point cloud restricted to different TV classes. While higher TV-classes show smaller standard deviations and deliver better overall accuracy, lower TV-classes are more likely to contain outliers (also cf. Fig. 2). TV classes greater than 8 are only present in flat areas facing the camera position. Since we focus on the reconstruction of vertical surfaces (i.e. facades) the information obtained by the test areas is extrapolated for all TV classes. The weighting function is derived by inverting the estimated function and defining the minimum weight with 1.0. Fig. 3. Box plots representing the standard deviation of the flatness error derived from different test areas for all available TV classes (top). Estimated weight function (bottom). PointCloudFusion.Thefusionof thepointcloudwascarried out in three iterationswithacylinder radiusof15cm(i.e. two times the GSD) and a height of 1.5 m. It is worth mentioning that, in some cases, during the image acquisition parts of the helicopter skids protruded into the camera angle, which leads to strong distortions in the matching procedure. The size of the octrees leaf node, which is used for the generation of the initialpointset, controls theapproximateoutputdensityof the fused point cloud. Therefore, faster runtimes can be achieved producing point clouds with lower density. The resolution used for the oblique imagery is set to 10 cm, to match the GSD of the input data. Within the point cloud fusion process, the points are filtered along the surface normal and weights are accumulated. The final surface representation is derived by discarding low weights, which are more likely to contain outliers. As depicted in Fig. 4, increasing the minimum weight threshold α leads to more accurate, however less dense point clouds. Fig. 4. Impact of rejecting low weighted points after the fusion procedure on density (top), accuracy and precision (bottom). Since the fusion method produces oriented point clouds, a mesh representation can be computed using Poisson surface reconstruction [8].The completeworkflowisdepicted inFig. 5. The runtime of the fusion process can be improved by discarding low level TV classes in a pre-processing step. However, the rejection of low level TV classes causes a loss in detail in areas with bad coverage. Evaluation. In order to measure the capability of the fu- sion routine different statistical measures are analyzed. The RMSE of the deviations between the reference point cloud and fused point cloud, give information about the accuracy of the 3D geometry. The standard deviation of the vertical digital surface model indicates the noise level of the point cloud, respectively the distribution of points perpendicular to the facade. As mentioned before, the density can be controlled by setting the octree resolution and by regulating the threshold for the minimum weightα. In Table I the raw point cloud is compared to the fused point cloud considering the influence of TV weights. The minimum weight threshold α is set to generate point clouds with comparable densities. Test areas include the school building located in the northern part of the mapped scene and the tower building located in the south. TABLE I COMPARISON OF THE FUSION ROUTINE REGARDING WEIGHTS. min. Density RMSE Fused Mean Fused Std. Dev. weightα [pnts/m2] PC-TLS [m] PC-TLS [m] of DSM [m] Raw (unfused) - 4398.00 0.199 0.108 0.296 Fused (no weights) 20 75.15 0.122 0.067 0.052 Fused (weighted) 30 74.25 0.111 0.063 0.040 Fused (weighted 18 75.23 0.102 0.049 0.032 pre-filter TV>1) 131