Exploiting Structural Regularities and Beyond: Vision-based Localization and Mapping in Man-Made Environments

Abstract

Image-based estimation of camera motion, known as visual odometry (VO), plays a very important role in many robotic applications such as control and navigation of unmanned mobile robots, especially when no external navigation reference signal is available. The core problem of VO is the estimation of the camera’s ego-motion (i.e. tracking) either between successive frames, namely relative pose estimation, or with respect to a global map, namely absolute pose estimation. This thesis aims to develop efficient, accurate and robust VO solutions by taking advantage of structural regularities in man-made environments, such as piece-wise planar structures, Manhattan World and more generally, contours and edges. Furthermore, to handle challenging scenarios that are beyond the limits of classical sensor based VO solutions, we investigate a recently emerging sensor — the event camera and study on event-based mapping — one of the key problems in the event-based VO/SLAM. The main achievements are summarized as follows. First, we revisit an old topic on relative pose estimation: accurately and robustly estimating the fundamental matrix given a collection of independently estimated homograhies. Three classical methods are reviewed and then we show a simple but nontrivial two-step normalization within the direct linear method that achieves similar performance to the less attractive and more computationally intensive hallucinated points based method. Second, an efficient 3D rotation estimation algorithm for depth cameras in piece-wise planar environments is presented. It shows that by using surface normal vectors as an input, planar modes in the corresponding density distribution function can be discovered and continuously tracked using efficient non-parametric estimation techniques. The relative rotation can be estimated by registering entire bundles of planar modes by using robust L1-norm minimization. Third, an efficient alternative to the iterative closest point algorithm for real-time tracking of modern depth cameras in ManhattanWorlds is developed. We exploit the common orthogonal structure of man-made environments in order to decouple the estimation of the rotation and the three degrees of freedom of the translation. The derived camera orientation is absolute and thus free of long-term drift, which in turn benefits the accuracy of the translation estimation as well. Fourth, we look into a more general structural regularity—edges. A real-time VO system that uses Canny edges is proposed for RGB-D cameras. Two novel alternatives to classical distance transforms are developed with great properties that significantly improve the classical Euclidean distance field based methods in terms of efficiency, accuracy and robustness. Finally, to deal with challenging scenarios that go beyond what standard RGB/RGB-D cameras can handle, we investigate the recently emerging event camera and focus on the problem of 3D reconstruction from data captured by a stereo event-camera rig moving in a static scene, such as in the context of stereo Simultaneous Localization and Mapping.