With the development of societies, the exploitation of mountains and forests is increasing to meet the needs of tourism, mineral resources, and environmental protection. The point cloud registration, 3D modeling, and deformation monitoring that are involved in surveying large scenes in the field have become a research focus for many scholars. At present, there are two major problems with outdoor terrestrial laser scanning (TLS) point cloud registration. First, compared with strong geometric conditions with obvious angle changes or symmetric structures, such as houses and roads, which are commonly found in cities and villages, outdoor TLS point cloud registration mostly collects data on weak geometric conditions with rough surfaces and irregular shapes, such as mountains, rocks, and forests. This makes the algorithm that set the geometric features as the main registration parameter invalid with uncontrollable alignment errors. Second, outdoor TLS point cloud registration is often characterized by its large scanning range of a single station and enormous point cloud data, which reduce the efficiency of point cloud registration. To address the above problems, we used the NARF + SIFT algorithm in this paper to extract key points with stronger expression, expanded the use of multi-view convolutional neural networks (MVCNN) in point cloud registration, and adopted GPU to accelerate the matrix calculation. The experimental results have demonstrated that this method has greatly improved registration efficiency while ensuring registration accuracy in the registration of point cloud data with weak geometric features.
Keywords: cosine similarity; multi-view convolutional neural networks; point cloud registration; weak geometric features.