Resumen de Real-time ekf-based structure from motion
Automated tridimensional scene and egomotion estimation from the only input of a set of images taken by a monocular camera has been a long-term aim in the Computer Vision community since the decade of the eighties. As the result of intense research, the so-called Structure from Motion (SfM) problem has experienced great advances.
Automated detection of salient features in natural images and the poste- rior establishment of correspondences along the rest of the images has been one of the rst milestones in this thread of research. Also, geometric relationships between pairs of images have been modeled based on projective geometry concepts including the most general cases; assuming no known information about the scene and including degenerate con gurations and uncalibrated settings. From these relations and a set of matches, relative transformations between pairs of camera poses and scene structure can be estimated. This initial solution can be re ned in a non-linear optimization stage -known as Bundle Adjustment (BA)- in order to obtain a globally consistent estimation.
Di erently from this main trend based on pairwise initial estimation fol- lowed by Bundle Adjustment, filtering-based SfM estimation have received considerably less attention and still lacks a thorough analysis of the main geometric issues tackled by SfM research along the latest three decades: projective modeling of the scene, estimation under degenerate con gurations and self-calibration. This thesis' main aim is to assess the efficacy, accu- racy and quality of the filtering-based SfM incorporating all the previously mentioned subjects. Real-time performance at 30 frames per second has been demonstrated for the contributions presented in this thesis, opening the path for real applications.
Speci cally, the main contributions in this thesis are 1) the introduction of the projective concepts of low-parallax points and points at in finity in a filtering framework and, based on that, the proposal of a drift-free real-time mosaicing algorithm; 2) a projective coding for 3D points based on explicit inverse depth parametrization, able to cope with low and high parallax con gurations in an undelayed and unifi ed manner; 3) the use of a camera-centered filtering scheme that reduces linearization errors; 4) an efficient 1-Point RANSAC that exploits prior knowledge from filtering; 5) a probabilistic approach to model selection -opposed to the standard one based on geometric reprojection error and regularization terms- well suited to a Bayesian filtering scheme; and 6) a Sum of Gaussians filter that allows accurate full camera self-calibration in a filtering framework.
