Global pose estimation and tracking for RGB-D localization and 3D mapping

Abstract

This thesis presents a detailed account of novel techniques for pose estimation by using both, color and depth information from RGB-D sensors. Since pose estimation simultaneously requires an environment map, 3D scene reconstruction will also be considered in this thesis. Localization and mapping has been extensively studied by the robotics and computer vision communities and it is widely employed in mobile robotics and autonomous systems for performing tasks such as tracking, dense 3D mapping and robust localization. The central challenge of pose estimation lies in how to relate sensor measurements to the state of position and orientation. When a variety of sensors, which provide different information about the same data points, are available, the challenge then becomes part of how to best fuse acquired information at different times. In order to develop an effective algorithm to deal with these problems, a novel registration method named Point-to-hyperplane Iterative Closest Point will be introduced, analysed, compared and applied to pose estimation and key-frame mapping. The proposed method allows to jointly minimize different metric errors as a single measurement vector with n-dimensions without requiring a scaling factor to tune their importance during the minimization process. Within the Point-to-hyperplane framework two main axes have been investigated. Firstly, the proposed method will be employed for performing visual odometry and 3D mapping. Based on actual experiments, it has been shown that the proposed method allows to accurately estimate the pose locally by increasing the domain of convergence and by speeding up the alignment. The invariance is mathematically proven and results in both, simulated and real environments, are provided. Secondly, a method is proposed for global localization for enabling place recognition and detection. This method involves using the point-to-hyperplane methods within a Branch-and-bound architecture to estimate the pose globally. Therefore, the proposed method has been combined with the Branch-and-bound algorithm to estimate the pose globally. Since Branch-and-bound strategies obtain rough alignments regardless of the initial position between frames, the Point-to-hyperplane can be used for refinement. It will be demonstrated that the bounds are better constrained when more dimensions are considered. This last approach is shown to be useful for solving mistracking problems and for obtaining globally consistent 3D maps. In a last part of the thesis and in order to demonstrate the proposed approaches and their performance, both visual SLAM and 3D mapping results are provided.