LiDAR Inpainting of UAV Based 3D Point Cloud Using Supervised Learning

Unmanned Aerial Vehicles (UAV) can quickly scan unknown environments to support a wide range of operations from intelligence gathering to search and rescue. LiDAR point clouds can give a detailed and accurate 3D representation of such unknown environments. However, LiDAR point clouds are often sparse and miss important information due to occlusions and limited sensor resolution. Several studies used inpainting techniques on LiDAR point clouds to complete the missing regions. However, these studies have three main limitations that hinder their use in UAV-based environment 3D reconstruction. First, existing studies focused only on synthetic data. Second, while the point clouds obtained from a UAV flying at moderate to high speeds can be severely distorted, none of the existing studies applied inpainting to UAV-based LiDAR point clouds. Third, all existing techniques considered inpainting isolated objects and did not generalise to inpainting complete environments. This paper aims to address these gaps by proposing an algorithm for inpainting point clouds of complete 3D environments obtained from a UAV. We use a supervised learning encoder-decoder model for point cloud inpainting and environment reconstruction. We tested the proposed approach for different LiDAR parameters and different environmental settings. The results demonstrate the ability of the system to inpaint the objects with a minimum average Chamfer Distance (CD) loss of 0.028 at a UAV speed of 5 ms \(^{-1}\) . We present the results of the 3D reconstruction for a few test environments.