Abstract
An autonomous robot able to navigate inside an unknown environment and reconstruct full 3D scene models using monocular video has been a long term goal in the field of Machine Vision. A key component of such a system is the reconstruction of surface models from estimated scene structure. Sparse 3D measurements of real scenes are readily estimated from N-view image sequences using structure-from-motion techniques. In this thesis we present a geometric theory for reconstruction of surface models from sparse 3D data captured from N camera views. Based on this theory we introduce a general N-view algorithm for reconstruction of 3D models of arbitrary scenes from sparse data. Using a hypothesise and verify strategy this algorithm reconstructs a surface model which interpolates the sparse data and is guaranteed to be consistent with the feature visibility in the N-views. To achieve efficient reconstruction independent of the number of views a simplified incremental algorithm is developed which integrates the feature visibility independently for each view. This approach is shown to converge to an approximation of the real scene structure and have a computational cost which is linear in the number of views. Surface hypothesis are generated based on a new incremental planar constrained Delaunay triangulation algorithm. We present a statistical geometric framework to explicitly consider noise inherent in estimates of 3D scene structure from any real vision system. This approach ensures that the reconstruction is reliable in the presence of noise and missing data. Results are presented for reconstruction of both real and synthetic scenes together with an evaluation of the reconstruction performance in the presence of noise.