Efficient Direct Voxel-Based Viewpoint Planning in 3D Reconstruction for UAVs

Abstract

This thesis presents a model-based path planning system for 3D reconstruction. Existing model-based methods generate surfaces using point cloud data from a first-phase path and then plan a second-phase path to cover the entire surface. However, these methods face two key challenges. First, the second- phase path often overlaps with the first-phase path, increasing flight time. Second, since viewpoints are sampled only for identified surfaces, unknown surfaces from the first phase may remain unobserved in the second phase. This study addresses these issues by tackling both the redundant path problem and the unknown surface problem. An unmanned aerial vehicle (UAV) is equipped with a LiDAR sensor to generate an occupancy grid map and a camera to collect images for offline 3D reconstruction. The proposed path planning system, like other model-based methods, consists of two phases. The first-phase path encircles a building, utilizing terrain elevation and building information from a topographic map. After the first-phase flight is completed, unexplored regions are identified using free and occupied voxels as well as unknown spaces. Viewpoints are then sampled directly from the voxel data using the proposed method. The second-phase planner connects these sampled viewpoints, eliminating redundant paths and effectively identifying previously unknown surfaces. This approach reduces flight time while slightly improving the coverage ratio.