Information-Driven Navigation

Kurzfassung

In the last years, we have witnessed an impressive progress in the accuracy and robustness of Visual Odometry (VO) and Simultaneous Localization and Mapping (SLAM). This boost in the performance has enabled the first commercial implementations related to augmented reality (AR), virtual reality (VR) and robotics. In this thesis, we developed new probabilistic methods to further improve the accuracy, robustness and efficiency of VO and SLAM. The contributions of our work are issued in three main publications and complemented with the release of SID-SLAM, the software containing all our contributions, and the challenging Minimal Texture dataset. Our first contribution is an information-theoretic approach to point selection for direct and/or feature-based RGB-D VO/SLAM. The aim is to select only the most informative measurements, in order to reduce the optimization problem with a minimal impact in the accuracy. Our experimental results show that our novel criteria allows us to reduce the number of tracked points down to only 24 of them, achieving state-of-the-art accuracy while reducing 10× the computational demand. Better uncertainty models for visual measurements will impact the accuracy of multi-view structure and motion and will lead to realistic uncertainty estimates of the VO/SLAM states. We derived a novel model for multi-view residual covariances based on perspective deformation, which has become a crucial element in our information-driven approach. Visual odometry and SLAM systems are typically divided in the literature into two categories, feature-based and direct methods, depending on the type of residuals that are minimized. We combined our two previous contributions in the formulation and implementation of SID-SLAM, the first full semi-direct RGB-D SLAM system that uses tightly and indistinctly features and direct methods within a complete information-driven pipeline. Moreover, we recorded Minimal Texture an RGB-D dataset with conceptually simple but challenging content, with accurate ground truth to facilitate state-of-the-art research on semi-direct SLAM.