Seite - 167 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Bilateral Filters for quick 2.5D Plane Segmentation Simon Schreiberhuber1, Thomas Mo¨rwald2 and Markus Vincze1 Abstract—We present a simple and practicable approach to segment organized point clouds gathered with RGBD sensors intoplanarelements.Thealgorithmproves toexecuteextremely fast while delivering all the dominant planes of a scene. As an integral part of our segmentation algorithm we examined two off the shelf and one heavily modified filtering algorithms to increase the quality of the point cloud before the actual segmentation process. The results of two of these algorithms were promising. One provides a favorable tradeoff between speed and quality while the other delivers superior quality at high computational cost. I. INTRODUCTION In mobile robotics many tasks have to be fulfilled in indoor environments. More specifically one task could e.g. include the search or classification of objects lying on the floor. Instead of processing all the points captured by the RGBD sensor it would be beneficial to early on discard some of the points that can not be part of the task. Removing the dominant planes from the scene is one common measure to achieve this. This becomes obvious when we observe that indoor environments are dominated by planar surfaces. While other plane segmentation algorithms operate on unfil- tered depth data, our algorithm utilizes a filtering step. Data as it is captured by an RGBD sensor tends to have multiple sources of noise, all of which tend to make the fitting of planes difficult. Reducing the noise upfront therefore is a prerequisite to a fast and simple plane segmentation approach. To create ideal conditions for our plane segmentation algo- rithm we discuss three filter approaches. With these filters we aim to refine planar regions while keeping the geometric details where they are needed. We show the results generated by the standard Bilateral Filter [7], the Sigma Adaptive Bilateral Filter [2] and the adapted Bilateral Mesh Denoising algorithm [4]. A discussion shows how these filters relate to each other and how they behave in specific situations. We describe the modifications necessary to apply the Bilateral Mesh Denoising algorithm to depth data and demonstrate its effectiveness. Regarding the core of our plane segmentation, we offer a comparison to two other algorithms: The comparably slow approach shown by Holz [5] which uses RANSAC to refine a rough normal based plane segmentation and an approach 1Simon Schreiberhuber and Markus Vincze are with the Vi- sion4Robotics group (ACIN - TU Wien), Austria{schreiberhuber, vincze}@acin.tuwien.ac.at 2Thomas Mo¨rwald was a member of the Vision4Robotics group. This work is supported by the European Comission through the Hori- zon 2020 Programme (H2020-ICT-2014-1, Grant agreement no: 645376), FLOBOT. shown by Wang [8] where a rough segmentation is improved on a point-wise basis. Both algorithms start with clustering the points into a 3D voxel grid. By doing this they are re- placing the inherent neighborhood information with a costly spacial relation. Finding the nearest neighbors to a specific point no longer is a simple access to the neighboring depth pixels but a search of all points in the adjacent voxel blocks. For our segmentation we follow a similar two-step approach as in [8] but make use of the neighborhood information contained in the organized point cloud. II. RELATED WORK Most plane segmentation approaches can be assigned to two categories. A direct approach, where planes are directly matched with the existing points, and indirect approaches where the scene is transformed into another representation. RANSAC [3] is a direct approach that iteratively tests randomly generated plane hypothesis against a point cloud and is often used to find the ground plane of a scene. To extractmultipleplanes fromasceneRANSAChas tobeused repeatedly to assign points to different planes. The outcome of this approach is highly dependent on the order in which the RANSAC algorithm finds the planes. Thus the affiliation of points to planes is ambiguous. The approach shown in [5] therefore does not use RANSAC for the segmentation itself but uses it to refine already existing plane hypothesis. These hypothesis are generated by clustering normal vectors in normal space or spherical coordinates. This delivers clusters of points, each of which is assembled by multiple planes facing the same direction. Averaging the normals within each of these clusters leads to a plane hypothesis which allows to separate the points into their according planes. Calculating the distance of the points to these plane hypothesis directly allows to cluster these points into their according planes. A more direct approach was chosen in [8] is based on roughly clustering plane patches within a 3D voxel grid. Some of these blocks within the voxel grid are containing enough points to approximate planes. In the following step it is possible to connect neighboring grid blocks to bigger surfaceswherever theseplanesare facing in roughly thesame direction. The approach chosen by Zhang [10] is to find lines along the horizontal scan-lines which are cuts trough planes. In a second step the normals get estimated along these line segments to find corresponding segments between scan-lines. Fitting line segments can then be connected to a planar region. The V-disparity algorithm [11] transforms the 3D data into a V-disparity map and therefore reduces the 3D plane fit to 167