Semantic labeling of lidar point clouds for UAV applications

Abstract

Small Unmanned Aerial Vehicle (UAV) platforms equipped with compact laser scanners provides a low-cost option for many applications, including surveillance, mapping, and reconnaissance. For these applications, semantic segmentation or semantic labeling of each point in the lidar point cloud, is important for scene-understanding. In this work, we evaluate methods for semantic segmentation of three-dimensional (3D) point clouds of outdoor scenes measured with a laser scanner mounted on a small UAV. We compare the performance of four different semantic segmentation methods, which are all applied in a scan-by-scan fashion, on semi-sparse laser data. The best method achieves 95.3% on the three classes ground, vegetation, and vehicle in terms of mean intersection over union (mIoU) on a previously unseen scene from a different geographical area. The results demonstrate that it is possible to achieve good performance on the semantic segmentation task on data measured using a combination of a small UAV and a compact laser scanner.