Reconstruction of a Digital Model of Objects Base on Aerial Laser Scanning Data

The article is devoted to the problem of restoring the geometry of objects based on airborne laser scanning data. The objective of the article is to develop a methodology and algorithms for restoring digital models of objects by generating missing point cloud data. The paper considers methods for processing point clouds, including segmentation using the PointNet++ neural network architecture, as well as restoring missing surface areas using the SPAR3D neural network model. Data was collected using a Velodyne VLP-32 lidar and a camera mounted on a UAV. The proposed approach consists of the stages of preprocessing, segmentation, mesh generation, point cloud supplementation, and point cloud stitching using the Fast Global Registration and Iterative Closest Point methods. The methodology was tested on a section of the Amur River, which made it possible to evaluate the efficiency of restoring objects located in the water protection zone. The completeness of the reconstructed surface of geometrically complex objects (boat, car) is estimated at at least 85% with a mean square positioning error of no more than 0.15 m relative to the original data. The methodology and algorithms can be adapted for other types of landscapes and infrastructure objects.