Watertight Surface Reconstruction for Uncertain Data

Kurzfassung

For simple pick and place tasks, a robot needs to be able to recognise objects. In order to avoid scanning all objects by hand, an automatic modelling approach is necessary. Considering space and price restrictions on a robotic platform, the quality of perceiving devices suffers from this restrictions. Consequently modelling approaches need to cope with pose errors and noisy data that usually result in displaced surfaces incorporating a lot of artefacts. In this work, we focus on creating a watertight surface representation that is automatically built by a robot. Therefore, we develop a next best viewpoint planning method that fits to an afterwards applied filtering stage for improving measurement confidence. Using the resulting data, a mesh growing approach reconstructs the surface of the to be scanned object incorporating an inflating and a detailing stage. Our approach is able to roughly estimate the objects surfaces using inaccurate hardware and even to reconstruct small details for precise laser scanning devices. Small areas that are not scanned can be filled and highlighted by utilizing the near neighbourhood of measurements.