Seite - 170 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

(a)mp. (b) cp Fig. 5: Curvaturemp and unsteadyness cp. These measures are used to guide σ in the adaptive filter and the normal estimation in the tangential filter. (a) Correct normals. (b) Reasonable result. (c) Poor normal calculation. (d) Poor filtering. (e) Even along plane surfaces, poorly calculated normals are leading to artifacts or instabili- ties. Fig. 6: Filtering results of the proposed Bilateral Tangential filter. slight bending towards fronto-parallelity. Note that this effect gets stronger as the planes get tilted. The Sigma Adaptive Bilateral Filter gets rid of this effect by not filtering in these critical regions. As seen in Figures 3c and 3d the results are comparable to the standard Bilateral Filter but without creating the ski defects. Spikes, as they often appear at sharp edges, will unfortunately not undergo any smoothing. Since we selected this filter to support our segmentationwecreatedanGPU(AMDRadeonHD6750M) implementation that computes within 25ms. For the Bilateral Mesh Denoising algorithm the results are different (see Figure 6). While being equally as suitable for the planar regions as the Bilateral and Sigma Adaptive Bilateral Filter, this algorithm shows the best results along discontinuities. In terms of computational complexity this al- gorithm unfortunately is way more demanding than the other two. This is mainly due to the complex normal estimation but also because it needs two to three iterations the other algorithms compute within one. V. SEGMENTATION Two examples for state of the art algorithms coming close to a 30Hz segmentation rate are [5] and [8]. The algorithm shown in [5] utilizes a segmentation in normal space but is slower than the other. We therefore follow the approach shownin[8]where thepointsaresplit intoa3Dvoxelgrid.A coarse pre-segmentationin on these then segments a majority of points with a relatively small amount of computations. Although the this approach is the faster one, it still is overly complicated for our needs. Seperating the organized pointcloud into equally sized cubes (voxels) only creates the need to compare these voxels to their 26 neighbouring voxels. A. Hierarchical Plane Segmentation The proposed algorithm follows the idea of pre-segmentation and splits the depth image into smaller patches similar to [8] but does it in image space. This reduces the number of neighbors for each patch to 8 and therefore saves computa- tion time. The main steps of the algorithms are: 1) Patch generation: The depth data is grouped into equally sized section with sizes like e.g. 10×10pixel. It is then tried to fit a plane into these points. If there are enough points within a threshold of this plane the patch is retained and the points will be assigned to this patch. When this criteria is not met the patch is discarded and the according points stay unassigned. 2) Patch Segmentation: The initially unassigned patches get grouped together to assemble planes. This happens according to their normal vector and position. 3) Post filtering: During this phase, no new patches will be added, but every pixel, which is bordering onto a plane and meets certain conditions, will be assigned to this plane. 1) Patches: As already mentioned. Patches are small equally sized fragments of the depth image and described by their plane equation: ax+by+cz−1=0 (17) The parameters can be acquired by principal component analysis of all points. After getting the parameters, it is necessary to test if they provide a good description of the plane. To ensure this, at least a certain percentage (e.g. 90%) of the points considered for this patch should be inside the 170