| Space resection is a fundamental task in many photogrammetric, remote sensing and computer vision applications. While in traditional photogrammetric techniques this task is exclusively solved from 3-D to 2-D point correspondences, recent advances in computer technology require us to adopt more general methods, accommodating features other than points, especially linear features. In these methods, image orientation is established based on a correspondence between a set of 3-D curves in object space and their observed 2-D image-space projections. Here, a prior correspondence between individual points describing the feature in object space and in image space is not required. In the last decade, substantial research on the use of higher level geometric features has been carried out in the photogrammetry and computer vision communities.; However, the types of modeled linear features have been rather limited. In this dissertation a hierarchical least squares method that solves for the exterior orientation parameters analytically from a set of general 3-D curves in object space and their 2-D images is derived. First, a non-linear least-squares method utilizing 3-D curves represented parametrically is developed. In the proposed method two types of local curve descriptors, independently extracted from object and image curves, are related. The proposed method is then extended to accommodate free-form space curves represented as an ordered list of 3-D points. Finally, a coarse-to-fine strategy is introduced to hierarchically recover the orientation parameters.; Although the correspondence between 3-D curves and their 2-D partial images is not solved in this work, one of its major contributions is the fact that given this correspondence, the association between their individual points is established as a by product from the resection solution. In that sense, the matching and the orientation tasks are carried out simultaneously. Furthermore, the beauty of the proposed method is that the underlying principal is not tied exclusively with the perspective transformation governing the image formation process. In fact, it is general enough to be employed for any other functional form relating measurable quantities in two arbitrary domains. Finally, the overall strategy laid out in this dissertation can be employed as a fundamental ingredient in a variety of computer vision and digital photogrammetry applications. |
