Efficient Structure and Motion: Path Planning, Uncertainty and Sparsity

This thesis explores methods for solving the structure-and-motion problem in computer vision, the recovery of three-dimensional data from a series of two-dimensional image projections. The first paper investigates an alternative state space parametrization for use with the Kalman filter approach to simultaneous localization and mapping, and shows it has superior convergence properties compared with the state-of-the-art. The second paper presents a continuous optimization method for mobile robot path planning, designed to minimize the uncertainty of the geometry reconstructed from images taken by the robot. Similar concepts are applied in the third paper to the problem of sequential 3D reconstruction from unordered image sequences, resulting in increased robustness, accuracy and a reduced need for costly bundle adjustment operations. In the final paper, a method for efficient solution of bundle adjustment problems based on a junction tree decomposition is presented, exploiting the sparseness patterns in typical structure-and-motion input data