Automatic matching and three-dimensional reconstruction of free-form linear features from stereo images

Automatic matching of free-form linear features in overlapping large-scale imagery over urban areas still remains to be a problem in both the photogrammetric and computer vision communities. Although there is a variety of algorithms that have been developed to solve this problem, reliable results are not always guaranteed. Differences in illumination conditions, relief displacement, and occlusions are some of the factors that make solving the matching problem more challenging. The deficiency of available techniques stems from the fact that they do not consider the perspective geometry of the imaging system in the matching process. Moreover, it is usually assumed that conjugate entities are almost exact copies of each other (this is rarely the case). The need for a reliable algorithm that can handle large-scale imagery over urban areas is growing, especially with the increasing availability of high-resolution aerial imagery. In this research, a new approach for automatic matching and three-dimensional reconstruction of free-form linear features from stereo images is proposed. The suggested strategy is based on The Modified Iterated Hough Transform (MIHT) for Robust Parameter Estimation. MIHT relies on the mathematical relationship between conjugate entities (the coplanarity condition when dealing with a stereo pair). As a result, it overcomes problems arising from relief displacements and/or occlusions. Moreover, MIHT simultaneously solves for the relative orientation parameters as well as the correspondence between conjugate entities in a stereo pair. Matching ambiguities along and across conjugate epipolar lines are observed in the correspondence output from MIHT. Ambiguities along the epipolar lines stem from the nature of the coplanarity model, which establishes the correspondence between conjugate epipolar lines rather than conjugate points. On the other hand, ambiguities across the epipolar lines are expected due to the discrete cell size of the implemented accumulator array in MIHT. In this work, the ambiguities ore resolved through several pruning techniques such as correlation coefficient, edge connectivity, height information, and epipolar geometry. Conjugate entities with sub-pixel accuracy are finally used to reconstruct accurate and reliable three-dimensional image-space discontinuities. Various experiments using large-scale imagery over urban areas proved the reliability and robustness of the proposed method.