Pose estimation techniques for monocular visual odometry and SLAM

Abstract

One of the key problems in computer vision and robotics is pose estimation as it is essential towards applications such as navigation, 3D reconstruction, Augmented Reality, and more. This thesis advances the state of the art in pose estimation techniques for monocular visual odometry and SLAM problems in terms of efficiency and robustness. Firstly, an iterative 5-point algorithm which solves the relative pose between two cameras viewing the same scene given 5 point correspondences is presented. This method provides a simple and efficient way for solving the relative pose between two cameras, and is found to provide more robustness towards noise. Next, the problem of visual tracking in a large environment is investigated. A monocular visual SLAM system which does not rely on global consistency is presented. The proposed system builds a sparse image graph using reference images that share some of their viewpoints of the operating environment. This allows the map to be optimized in a very efficient manner at the expense of global consistency. The pose of the camera is then localized with respect to a reference image in the map using 2D point correspondences with respect to a reference image in the map using 2D point correspondences between the live view and several reference images. The system is locally accurate, very efficient compared to state of the art visual SLAM systems for large environments, and can easily enable applications to be built on top of it. Finally, this thesis investigates the problem of pose estimation using whole image alignment techniques. A new technique for performing an efficient variant of image alignment which handles missing data and re-weighting is presented. Using a preconditioning strategy, the proposed method does not require the Jacobian and Hessian matrix to be re-computed at every iteration. The proposed method is found to be more efficient and equally robust compared to the state of the art image alignment techniques.