Efficient 3D Segmentation, Registration and Mapping for Mobile Robots

Abstract

Sometimes simple is better! For certain situations and tasks, simple but robust methods can achieve the same or better results in the same or less time than related sophisticated approaches. In the context of robots operating in real-world environments, key challenges are perceiving objects of interest and obstacles as well as building maps of the environment and localizing therein. The goal of this thesis is to carefully analyze such problem formulations, to deduce valid assumptions and simplifications, and to develop simple solutions that are both robust and fast. All approaches make use of sensors capturing 3D information, such as consumer RGBD cameras. Comparative evaluations show the performance of the developed approaches.<br /> For identifying objects and regions of interest in manipulation tasks, a real-time object segmentation pipeline is proposed. It exploits several common assumptions of manipulation tasks such as objects being on horizontal support surfaces (and well separated). It achieves real-time performance by using particularly efficient approximations in the individual processing steps, subsampling the input data where possible, and processing only relevant subsets of the data. The resulting pipeline segments 3D input data with up to 30Hz.<br /> In order to obtain complete segmentations of the 3D input data, a second pipeline is proposed that approximates the sampled surface, smooths the underlying data, and segments the smoothed surface into coherent regions belonging to the same geometric primitive. It uses different primitive models and can reliably segment input data into planes, cylinders and spheres. A thorough comparative evaluation shows state-of-the-art performance while computing such segmentations in near real-time.<br /> The second part of the thesis addresses the registration of 3D input data, i.e., consistently aligning input captured from different view poses. Several methods are presented for different types of input data. For the particular application of mapping with micro aerial vehicles where the 3D input data is particularly sparse, a pipeline is proposed that uses the same approximate surface reconstruction to exploit the measurement topology and a surface-to-surface registration algorithm that robustly aligns the data. Optimization of the resulting graph of determined view poses then yields globally consistent 3D maps. For sequences of RGBD data this pipeline is extended to include additional subsampling steps and an initial alignment of the data in local windows in the pose graph. In both cases, comparative evaluations show a robust and fast alignment of the input data.