Real-time scalable structure from motion: from fundamental geometric vision to collaborative mapping

Good egomotion estimation forms the backbone of any modern highperformance localization and navigation system. In contrast to global positioning systems or laser-based range measurement devices, the use of cameras represents an increasingly interesting alternative promising the applicability in a vast number of related scenarios. The range of possible fields of application easily extends from indoor to outdoor, small-scale to large-scale, and under-water to aerial operations. The computer vision community has investigated camera-based egomotion estimation for more than three decades now—research that has lead to impressive results. Two fundamental ways have been pursued. The first one consists of a purely geometric approach, where the incremental transformation between consecutive images is each time computed by absolute or relative camera pose algorithms, based on the identified feature correspondences. The second approach additionally takes time information into account, and estimates priors about the relative camera displacement by means of a motion model. Each modality can be extended by taking additional sensorial information into account, such as inertial readings from an IMU, GPS data, or laser range information. The present dissertation represents a contribution towards purely geometric visual egomotion-estimation approaches, which are of major importance for initializing model-based solutions or robustifying estimation of motion with challenging dynamics. The scope ranges from fundamental algebraic geometry to the practical realization of systems for real-time motion estimation on real-life image sequences. The main cornerstones of this dissertation are given by a number of novel geometrical solutions for absolute and relative camera-pose computation in the calibrated case. The presented minimal solution for absolute camera-pose computation sets a new standard in terms of efficiency and numerical robustness. This algorithm is then extended to the non-perspective case via minimal and linear complexity n-point solutions. Finally, the derivation of an intuitive novel epipolar constraint leads to a minimal solution for direct translation-independent computation of the