With the development of optical photogrammetry technologies, there are more and more means to perceive three-dimensional (3D) point clouds that describe the same object or scene. Through satellite photogrammetry, we can quickly obtain dense urban point clouds in a wide area, but the surface information of the target is not clear because the sensor is too far away, and even the data collected by the close-range platform has fine structure and texture information. However, when there is no precise positioning system or absolute control points, the generated point cloud is in an arbitrary model coordinate system. The rapid and high-precision registration between large-scale point clouds of close-range images and point clouds of satellite images has great potential for applications such as smart city construction, disaster relief, and emergency response. However, there are many problems in this task, which makes it difficult to achieve efficient registration between the two. For example, the resolution of satellite images and close-range images is different, which leads to a large difference in the point density between the two point clouds. As the sensor's line of sight is blocked, there are many holes in point clouds of images. The scale difference in the coordinate system between the close-range point cloud and the satellite point cloud is arbitrary. The image point cloud contains a large number of noise points and outliers because of defects in the dense image-matching algorithm. To this end, an efficient cross-source image point cloud registration method is proposed on the basis of graph theory, which is automatic, fast, and robust. It is believed that the proposed basic registration strategy and graph-matching method can be helpful for the data fusion and reconstruction of large-scale satellite image point clouds and close-range image point clouds.

First of all, the ground plane direction in the point cloud is found through the geometric features of the point cloud. The rotation angle of the planar normal vector with respect to the vertical direction of the satellite point cloud is calculated so that the close-range point cloud is roughly aligned with the satellite point cloud on the ground. Then, the centers of the buildings are taken as the nodes, and the layout relationship of buildings in the point cloud is constructed into a graph, which transforms the point cloud registration problem into a graph-matching problem. Afterward, kernel triangles are constructed according to geometric constraints as registration primitives, and higher-order similarity information is used to find the global optimal match of graphs. Finally, the ICP algorithm is adopted for fine registration to obtain high-precision and cross-source point cloud registration results.

The point cloud of the Gaofen-7 satellite images and the point cloud of UAV close-range images in three regions of Henan Province are selected for experiments to verify the effectiveness of the proposed method. There are 22 pairs, 13 pairs, and 11 pairs of nodes that are matched in the three experiments (Fig. 3). In the three experiments with different numbers of holes and noise points, the graph with higher-order similarity information can accurately obtain a sufficient number of matching nodes, overcoming the density and scale differences. Upon the application of the ICP algorithm, the integrated point clouds are obtained, which not only show rich geometric structure and texture details but also have real geographic coordinates (Fig. 4). The coarse registration algorithm based on graph matching enables the ICP algorithm to avoid falling into a local optimal solution, which has good registration accuracy on three datasets of different scales, densities, and noise. The root-mean-square errors of the three experiments are only 5.16 m, 6.39 m, and 9.02 m (Table 2). Finally, the existing four algorithms are used to register three experimental datasets in this paper (Fig. 5) for the performance comparison with the proposed registration method. The experimental results show that the proposed method is independent of noise points and outliers. It can overcome the density differences of different point clouds and eliminate coordinate scale differences of about 939 times. The overall registration speed is improved by a factor of 51-184 compared to that of the comparison methods, and the proposed method is automatic, robust, and efficient.

This paper mainly studies the scale differences, density differences, noise points, and outlier problems in the registration of satellite-image point clouds and close-range image point clouds. A novel point cloud registration method is proposed, which transforms the point cloud registration problem into a graph-matching problem according to graph theory. The centers of buildings are taken as the nodes, and the layout relationship of buildings in the point cloud is constructed into a graph. Kernel triangles are constructed pursuant to geometric constraints as registration primitives. Then, a graph-matching method using the higher-order similarity information of the graph is presented to obtain the spatial transformation model, and the ICP algorithm is used for fine registration. Finally, experiments are conducted on high-resolution satellite-image point clouds and close-range image point clouds in three different regions of Henan Province. The close-range point cloud containing structure and texture details is successfully converted to the spatial coordinate system of the satellite point cloud, and a refined 3D point cloud with real geographic coordinates is obtained. Multiple sets of experimental data show that the proposed method can robustly and quickly register cross-source image point clouds in contrast with other methods.