| 英文摘要 | As one of the most important research directions in computer vision, 3D reconstruction is of both high academic value as well as wide applicability. Although, fruitful achievements have been obtained in the field, many core problems remain open to be solved. This thesis is focused on some key aspects of 3D reconstruction, such as camera self-calibration, reconstruction methods, image registration, single view metrology, the main work can be summarized as: 1. The robustness aspect of two main kinds of camera self-calibration methods, namely the absolute conic based method and the absolute quadric based one, is investigated and compared in term of the conditioning number of the corresponding coefficient matrix. Simulation results show that the absolute conic based method is more robust than the absolute quadric based one, a conclusion of being in direct contradiction with the previous reported observations in the literature. 2. A comparative study is done on two reconstruction methods, namely the essential matrix based method and the absolute dual quadric based method. Based on extensive experiments on simulated data as well as on real images, we shown that surprisingly enough, the absolute dual quadric based method consistently outperforms the essential matrix based one in terms of reconstruction accuracy and robustness, hence it is highly recommended to use the absolute dual quadric based method rather than the commonly used essential matrix based one in real application. 3. We investigate the theoretical basis of the polynomials based image As one of the most important research directions in computer vision, 3D reconstruction is of both high academic value as well as wide applicability. Although, fruitful achievements have been obtained in the field, many core problems remain open to be solved. This thesis is focused on some key aspects of 3D reconstruction, such as camera self-calibration, reconstruction methods, image registration, single view metrology, the main work can be summarized as: 1. The robustness aspect of two main kinds of camera self-calibration methods, namely the absolute conic based method and the absolute quadric based one, is investigated and compared in term of the conditioning number of the corresponding coefficient matrix. Simulation results show that the absolute conic based method is more robust than the absolute quadric based one, a conclusion of being in direct contradiction with the previous reported observations in the literature. 2. A comparative study is done on two reconstruction methods, namely the essential matrix based method and the absolute dual quadric based method. Based on extensive experiments on simulated data as well as on real images, we shown that surprisingly enough, the absolute dual quadric based method consistently outperforms the essential matrix based one in terms of reconstruction accuracy and robustness, hence it is highly recomme |
修改评论