| dc.description.abstract | This thesis develops new stereo vision methods for the reconstruction and analysis of complex surfaces, such as riverbeds. Depth and surface orientation estimates are crucial to the understanding of 3D scene geometry, from calibrated stereo images. In this work, we propose new visibility and disparity magnitude constraints, for slanted patches in the scene. These constraints can be used to associate geometrically feasible planes with each point, in a 3D disparity space representation. The new constraints are validated in the PatchMatch Stereo framework of Bleyer et al. (BMVC, 2011). In order to estimate the plane parameters in this algorithm, we modify the original spatial propagation procedure, and introduce a gradient-free non-linear optimiser. These improvements allow us to achieve accurate disparity maps, with sub-pixel precision, according to the Middlebury stereo benchmark. In addition to surface orientation, curvature information is needed for a full understanding of the local surface structure. The PatchMatch surface model is planar, and does not directly estimate the local curvature. We propose a local quadric surface model, which uses both the spatial position and the estimated surface normals, in the disparity space. We also propose principal curvature and principal direction constraints, which ensure that the local quadric model is geometrically feasible. Finally, we describe the design and capture of a new photogrammetric dataset, which can be used to study topographic changes in riverbed morphology, over time. The dataset is challenging for conventional stereo matching algorithms, because the visible surface consists of sand, which lacks large-scale image features. This laboratory dataset comprises thirty-nine calibrated stereo pairs, plus fifteen ground-truth depth maps, obtained by a laser scanner. We used this dataset to validate the stereo vision methods developed in the thesis, in relation to potential geomorphological applications. | en_US |
