Towards an Accurate RGB-D Benchmark, Mapping and Odometry as well as their Applications

Towards an Accurate RGB-D Benchmark, Mapping and Odometry as well as their Applications
Oliver Wasenmüller
PhD Thesis

Abstract:

RGB-D cameras have the unique capability to capture color as well as dense depth images at the same time. The advantage of depth images is that they describe the 3D surface of the scene, which makes them ideal candidates for reconstructing geometries. However, they have the disadvantage that the depth images have a considerable amount of noise which must be explicitly taken into account in subsequent algorithms. In this thesis, we investigate 3D reconstruction with these devices. The reconstruction problem can be divided into two main parts: the mapping – the fusion of several depth images to a consistent 3D model – and the odometry – the estimation of the camera pose over time. We propose a new benchmark - called CoRBS – allowing to independently evaluate such algorithms for the first time. Next, we present several new mapping algorithms outperforming the state-of-the-art in accuracy. In addition, we propose a new odometry algorithm designed for ToF cameras. At the end, we examine different application scenarios – like a 3D discrepancy check - for the presented algorithms.