混合式从运动恢复结构的理论和方法研究

Structure-from-Motion (SfM) is a fundamental problem in the scene 3D reconstruction field. It has a broad application prospects in the protection of intangible cultural heritage of ancient architecture, 3D city modeling and military maneuver. Currently, incremental SfM method is robust while its efficiency does not satisfy the demand of large-scale scene reconstruction. Though global SfM method is fast, it is sentive to the epipolar geometry outliers. Single reconstruction mode is successful in handling images with dense connections and few outliers, however, for large-scale scene, the sub-scenes are usually sparsely connected and contaminated by many outliers, current SfM methods cannot satisfy the demand of both fastness and robustness in the large-scale scene reconstruction. To tackle these problems, this research proposes a hybrid SfM framework, in which we are focusing the three following key problems: epipolar geometry graph constuction, scene grouping and merging, and camera poses estimation: 1) For the epipolar geometry graph construction problem, we fuse the scene structure, matching points distribution and detection accuracy to compute accurate epipolar geometries; 2) For the scene grouping and merging, we explore a probability distribution model for the epipolar geometry connections; 3) For the camera poses estimation, a novel hybrid estimation algorithm is studied to inherit the advantages of both incremental manner and global manner. This research is an enrichment in both theory and algorithms of large-scale scene reconstruction, which could promote the achievements in the practical application.

从运动恢复结构是场景三维重建的基础问题，其成果在古代建筑非物质文化遗产保护、城市三维建模、军事演习等领域有着广泛应用前景。现有增量式算法鲁棒但效率不足，全局式算法快速但对外点敏感。单一重建模式在图像连接紧密且外点少时比较有效，但大场景中子场景之间通常连接不够紧密、外点较多，现有算法无法满足快速鲁棒的重建需求。为克服现有算法不足，本项目拟构建一套混合式的从运动恢复结构框架，对框架中的外极几何图构建、场景分组与合并以及摄像机位姿求取等三个核心问题进行研究：1）针对外极几何图构建，研究结合场景结构、匹配点分布与定位精度的外极几何关系准确求取算法；2）针对场景分组与合并，研究基于外极几何图连接关系概率分布模型的处理机制；3）针对摄像机位姿求取，融合全局式和增量式两种模式的优势构造一套混合式的摄像机位姿求取方法。本项目研究是对大场景三维重建领域理论与方法的丰富，有助于推动这一领域成果的实际应用。

从运动恢复结构是场景三维重建的基础问题，其成果在古代建筑非物质文化遗产保护、城市三维建模、军事演习、增强现实等领域有着广泛应用前景。现有增量式算法鲁棒但效率不足，全局式算法快速但对外点敏感。单一重建模式在图像连接紧密且外点少时比较有效，但大场景中子场景之间通常连接不够紧密、外点较多，现有算法无法满足快速鲁棒的重建需求。为克服现有算法不足，本项目构建了一套混合式的从运动恢复结构框架，对框架中的外极几何图构建、特征点轨迹选择和混合式摄像机位姿鲁棒估计等三方面开展了系统性的研究，提出了增量式的外极几何图构建方法、基于正交最大生成树的渐进式相机位姿估计方法、基于投票策略和外极几何图连接关系的摄像机增量注册模式、快速鲁棒的大场景旋转平均估计策略和基于外极几何图覆盖的特征点轨迹选择等一系列的核心算法。本项目研究是对大场景三维重建领域理论与方法的丰富，有助于推动这一领域成果的实际应用。.