| dc.description.abstract | This study developed algorithms to measure and reconstruct three-dimentional (3D) objects and scenes from a single-perspective image based on its geometric cues.
Single view metrology relies on the 3 vanishing points that converge by parallel lines along 3 mutually orthogonal axes.
Vanishing points can be used to estimate not only the camera pose but also internal parameters; thus, the 3D measurement of monocular vision can obtain a partial or complete 3D reconstruction of a scene.
Advantages of the proposed method include low cost, reliable accuracy and flexibility, and the potential to reconstruct buildings and other architecture that are heavily damaged or even no longer exist.
The presented algorithms employ uncalibrated images; therefore, no prior camera information is needed.
Exterior and interior orientation parameters can be calibrated directly from the vanishing points.
The proposed scheme began with line segment extraction and classification using cascade Hough transform from photographs or paintings with fine-perspective projection, and a fully automated base point searching algorithm is then used to locate the projection of feature points on the referenced plane.
The systematic and random errors of the vanishing points and base points are minimized iteratively during the vanishing point refinement process based on the diversity of each projection group with an O(1) computational complexity.
Three-dimensional coordinates of feature points are computed based on the single-view metrology.
In addition, 3 types of non-planar structures including elliptic cylinder shapes, surfaces of revolution, and free-form surfaces are extracted separately for reconstruction based on their significant parameters.
Finally, regular and curved models are merged based on their shared feature points.
The output imagery of a computer-simulated model, video frame cut, consumer camera, and real paintings are used in this study to test the performance of the algorithms.
The proposed vanishing point refinement process is able to reduce the differences between the images with and without lens distortion.
Quantitative evaluations of the results compared with ground-based surveying and visualized comparison with raw images indicate that the algorithms can successfully extract 3D information and reconstruct 3D models of specific non-planar structures.
Estimation errors of regular and non-planar parameters are less than 1\% compared to the ground truth using computer simulated imagery.
For close range photograph, the average regular and non-planar errors are less than 2\% and 3\%, respectively, compared to the ground truth measured by ground-based LIDAR and stereo photo pairs.
The proposed vanishing point refinement process improves the RMSE of the 3D model from about 2\% error to about 0.6\%.
The accuracy of the proposed methods are related to the viewing angles based on the vanishing point geometry, and validation of viewing angle tolerance is also provided by given random errors during vanishing point calculation. | en_US |