A Fast Reconstruction Method Based on Real-time Terrain Following

Carefully planned flight trajectories for Unmanned Aerial Vehicle (UAV) are used to reconstruct high-quality 3D models of real environments, which greatly simplify the process of large-scale scene reconstruction and reduce the computing cost. However, in order to better capture the image of uneven terrain in the strange environment, most state-of-the-art methods need a prior map of the scene, and thus additional pre-processing steps are required before the actual image acquisition. To address this limitation and make unknown landscapes easier to be reconstructed, we propose an algorithm for real-time exploration and reconstruction that does not demand any prior map of the area. And it can quickly generate dense point clouds for the area. The algorithm uses the sparse point cloud generated by Simultaneous Localization and Mapping (SLAM) to guide the UAV movement, in addition it generate pruned dense 3D point clouds by constructing virtual stereo-pair from consecutive frame. Experiments show that the proposed method could generate effective flight trajectories that satisfy to the task, therefore, the 3D scene could be reconstructed quickly with comparable accuracy without prior map information.