LiDAR Inertial Mapping of Non-Manhattan Topologies for UAV-Based Aircraft Inspection in GNSS-Denied Environments

Abstract

In this paper we propose and evaluate a high resolution intensity yielding adaption of a 3D mapping approach for airplane inspection by autonomous unmanned aerial vehicles. We solely use a 16 column laserscanner and an industrial-grade inertial measurement unit, without a priori knowledge of the inspection object dimensions or its surroundings. Our approach includes the introduction of a novel cylinder feature detection and an according modification of the optimization-based state estimation. Robustness of the point cloud registration is proven based on measurements in a GNSS-denied hangar environment with a Boeing 737-500, comprising only sparse plane and edge features and for highly dynamic movement involving rapid roll or pitch rotation. The approach is suitable for the given application to datasets with strong non-Manhattan characteristics, as opposed to several state of the art algorithms. By means of including the intensity information, 2D projections of the 3D registration result depict a remarkable level of detail and sharpness, resembling that of camera images. Thus allowing for complete coverage of the environment with high spatial resolution using the discrete points and voxel grids with small leaf sizes, the obtained results perspectively can be used for visually matching multi-channel camera images. By fusion of such visual information with the LiDAR generated projections and using the LiDAR inertial odometry, 3D defect maps of the aircraft surface can be constructed, enabling the automated localization of structural damages that are detectable using convolutional neural networks, as shown in previous publications.