精准、高效重建大尺度城市场景三维几何信息是摄影测量与计算机视觉领域的研究热点。相比LiDAR技术，多视立体影像三维点云重建具有操作便捷、成本低、且带有光学纹理信息等优势，成为了城市场景三维几何信息采集的重要手段，但面临众多挑战亟待解决，如重建效率低、细节特征难保持以及弱纹理、重复纹理、遮挡区域深度重建鲁棒性差。因此本项目以多源、多平台影像为数据源，以实现大尺度城市场景三维几何信息高精度重建为研究目标；构建倾斜深度平面约束的深度重建能量模型，研究树动态规划层次块匹配能量最优化框架，实现快速、高精度深度图重建；构建图像引导深度增强能量模型，发展半全局块匹配能量最优化方法，实现误匹配深度及物体细节高精度修复；研究内外存调度的哈希格网与自适应八叉树双层索引深度图融合，实现多视深度图低冗余、高保真及高精度融合。本项目研究成果有望提高三维几何信息重建效率及精度，为智慧城市、智慧地球建设提供基础空间数据。

Accurate and efficient reconstruction of the larger-scale 3D geometric information of the urban scene has been one of the hot issues in the fields of photogrammetry and computer vision. Compared to the LiDAR technology，3D point cloud reconstruction of multi-view stereo has advantages of easy handling, low cost, and simultaneous acquisition of the optical texture information. It has become an important means of the 3D geometric information collection of urban scene, However, there are still several problems that need to be urgently solved, such as, low efficiency, difficulty to maintain the details of objects, and poor robustness of depth reconstruction at the noise, poor texture, repeated texture, and occluded area. Therefore, this item aims to reconstruct the high-accuracy 3D geometric information using the high-resolution images from multi-source and -platforms. Firstly, an energy model for depth reconstruction with slanted depth plane constraints is proposed, and a PatchMatch-based hierarchical energy optimization framework based on tree dynamic programming is construct to achieve the fast and accurate depth reconstruction of a large-scale urban scene. Secondly, we build an energy model of image guided depth enhancement, and an energy optimization method based on semi-global matching (SGM) and PatchMatch framework is developed to obtain the high-accuracy depth information of mismatched pixels and details. And finally, a depth fusion method based on out-of-core technology is created to achieve the fusion of multiple depth maps with a low redundancy, high fidelity and high accuracy by integrating the grid based hash and adaptive octree data structure. The results of the proposed approaches are expected to improve the efficiency and accuracy of 3D geometric information reconstruction, and provide basic spatial data for the construction of smart cities and smart planets.