3D scene reconstruction from optical remote sensing images is an important research topic in computer vision and remote sensing, which aims to automatically recover the 3D scene structure from a set of multi-view remote sensing images. In recent years, 3D reconstruction from optical remote sensing images has shown great application potential in many fields, such as urban planning, autonomous navigation, etc. This
thesis investigates the problems of low speed and completeness in 3D reconstruction from optical remote sensing images, and its main works include:
(1) Most of the existing methods in literature employ the RPC (Rational Polynomial Coefficients) model for 3D reconstruction from optical remote sensing images in an incremental reconstruction manner, and their reconstruction speed and completeness are generally low. To address this problem, a fast reconstruction method for optical satellite images based on the affine imaging model is proposed. Firstly, the input multi-view satellite images are cropped into a set of small-sized patches with overlapping regions, and for each pair of patches that have a sufficient number of point correspondences fromtwo views, the corresponding 3D affine point cloud is calculated. Then based on the
obtained local point clouds, a global affine camera motion estimation algorithm is explored for calculating the affine motion matrices of the cameras corresponding to all the patches in a unified coordinate system. Finally, the obtained affine camera motion matrices and a small number of ground control points are utilized to recover the Euclidean scene structure. The proposed method, which is independent of the prior information of the RPC model, does not perform bundle adjustment repeatedly, achieving a fast reconstruction for satellite images. Experimental results on the MVS3DM and DFC2019 datasets demonstrate that the reconstruction speed, accuracy and completeness of this method are better than three state-of-the-art 3D reconstruction methods in most cases.
(2) In order to further reduce the computational cost of 3D reconstruction from optical remote sensing images, a fast 3D reconstruction method based on scene graph partition is proposed. Specifically, we first construct an initial epipolar geometry graph according to the geometric information between pairs of images, and divide it into several independent sub-scene graphs by utilizing a clustering algorithm. Then, the maximum spanning tree is constructed according to the initial epipolar geometry graph, and it is used to expand the sub-scene graphs to enhance the connections among the obtained sub-scene graphs. Next, based on each expanded sub-scene graph, the corresponding local scene is reconstructed respectively. Finally, all the reconstructed local scenes are
merged into a complete 3D scene by utilizing the maximum spanning tree. Experimental results on two satellite optical remote sensing datasets and three UAV(Unmanned Aerial Vehicle) optical remote sensing datasets demonstrate that the proposed method achieves close reconstruction accuracy and completeness, but performs faster in comparison to four state-of-the-art methods.