基于多视角航拍图像的大规模城市场景快速重建

The digital technology for urban space reconstruction is one of the most popular research issues in computer vision and graphics domain. It has so many important applications such as urban planning and management, military simulation, digital entertainment and etc. Due to the lack of hardware algorithms and other factors, the most of current technologies cannot provide satisfactory reconstruction results. This project focuses on using of aerial images to quickly reconstruct large-scale urban scenes, especially those with higher density and highly building occlusion area. The three main research contents comprise: (1) large-scale three-dimensional urban point cloud reconstruction, (2) urban point cloud data classification and building rooftop segmentation, (3) building objects mesh reconstruction in large-scale three-dimensional urban scene. The key scientific issues comprise: depth estimation in sub-pixel level and multiple view depth information fusion, three-dimensional information enhanced building objects accurate segmentation, multi-feature fusion based building objects mesh reconstruction. Innovation of this research program is: (1) patchmatch based high accuracy depth estimation and multiple view depth information fusion, (2) three-dimensional point cloud semantic information enhanced building objects accurate segmentation, (3) fusion of space voxel and surface texture based building objects mesh reconstruction. The research project may directly provide services to construct the three-dimensional Geographic Information System (GIS) for the purpose of urban planning and management.

城市空间数字化技术是计算机视觉和图形学的热点研究问题，在城市规划与管理、军事仿真及数字娱乐等领域具有重要的应用价值。现有技术受到硬件条件和算法不足等因素制约，无法得到令人满意的重建结果。本项目专注于利用航拍图像来快速重建大规模城市场景，尤其是针对大面积具有较高密度和存在高度遮挡的建筑场景。三个研究内容为：(1)大规模三维城市场景点云重建；(2)城市场景点云数据分类及建筑对象分割；(3)大规模三维城市场景建筑对象网格重建。关键科学问题为：亚像素级别的深度估计及多视角深度信息融合、三维信息增强的建筑对象精细分割、多特征融合的建筑对象网格重建。研究方案的创新为：(1)基于片元匹配的高精度深度估计及多视角深度信息融合；(2)基于三维点云语义信息增强的建筑对象精细分割；(3)基于空间体素和表面纹理融合的建筑对象网格重建。本项目的研究成果可直接服务于城市规划及管理的三维地理信息系统（GIS）构建。

城市空间数字化技术是计算机视觉和图形学的热点研究问题，在城市规划与管理、军事仿真及数字娱乐等领域具有重要的应用价值。现有技术受到硬件条件和算法不足等因素制约，无法得到令人满意的重建结果。本项目专注于利用航拍图像来快速重建大规模城市场景，尤其是针对大面积具有较高密度和存在高度遮挡的建筑场景。..在大规模三维城市场景点云重建方面，提出了面向点云重建的多视角图像非对称哈希检索加速方法、面向点云重建的多视角图像运动模糊去除方法、基于多视图几何一致性约束的无监督稠密深度预测方法、通用的SLAM重建框架和基准、深度插值与物体分割的联合优化。..在城市场景点云数据分类及建筑对象分割方面，提出了基于分层RGB-D先验的多视角航拍图像建筑屋顶提取方法、基于显著特性的遥感图像建筑屋顶提取方法、基于局部形状的三维点云扰动的分类增强方法、融合多策略的建筑物图像语义解析技术框架、基于自适应先验选取策略的双目图像交互分割、基于多尺度特征融合的目标检测方法。..在大规模三维城市场景建筑对象网格重建方面，提出了基于多视角图像的过程式植物建模、基于多视几何及GPS约束的三维地图建模、宽基线航拍图像的移动端实时正射影像地图建模。..本项目发表论文19篇，其中SCI论文12篇，授权发明专利2项，另外申请发明专利1项，共计培养博士生5名，硕士生3名，博士后1名，并有一名学生晋升为高级职称。完成了预期的任务。