Out-of-Core Bundle Adjustment for 3D Workpiece Reconstruction

In this thesis, we developed an RGB-D-based 3D reconstruction framework with a special emphasis on out-of-core bundle adjustment. Our approach is suitable for 3D workpiece reconstruction as well as for the reconstruction of arbitrary scenes. We first acquire RGB-D data of the scene by using a hand-held RGB-D sensor. The acquired depth map is preprocessed to achieve reduced sensor noise. Camera tracking is performed using a robust feature-based 3D alignment algorithm based on RANSAC. An initial map of the environment, which consists of camera poses, 3D landmarks and observations, is built using frame-to-frame tracking and is augmented with detected loop closures. In order to reduce the effects of accumulated drift and inaccuracies in the map over time, bundle adjustment is applied. We minimize 3D alignment errors instead of 2D reprojection errors to improve robustness, convergence behaviour and accuracy. Since full bundle adjustment is unfeasible for large datasets, we introduce a novel combination of an efficient submap-based approach together with the minimization of 3D alignment errors. Our submap-based approach partitions the bundle adjustment problem into submaps, which are optimized independently of each other afterwards. After an efficient global alignment of the submaps, they are again optimized internally, but with fixed separators. We finally integrate the RGB-D frames into an octree-based 3D representation to generate a dense 3D model. Our submap-based bundle adjustment method significantly improves the runtime compared to full bundle adjustment, especially for large datasets. Further, our method approaches the accuracy of full bundle adjustment and outperforms the RGB-D SLAM system. Finally, we present reconstructed 3D models of workpieces of compelling visual quality and metric accuracy, which makes them suitable for visual inspection, measuring tasks and reverse-engineering.