3D-Surface Reconstruction

点击翻译按钮获取中文摘要

The focus of P5 is the estimation of dense and high-resolution 3D reconstruction from the available image data. To this end, we pursue a two-stage strategy:(1) Based on the onboard stereo camera, we aim at computing a crude semi-dense visual SLAM (depth maps, geometry and camera motion). It is computed at video frame rate on the quadrotor and takes into account the statistical uncertainty of 3D locations and image information.(2) On the ground station, we improve the semi-dense 3D reconstruction toward a high-resolution dense 3D reconstruction by also taking into account the high-resolution images from the main sensor. More specifically we plan to generate dense depth maps which are then fused to a dense 3D model. This step will also take into account 3D point clouds from bundle adjustment and 3D point clouds from the laser scanner. In this second phase of the project, we will emphasize robustness to perturbations through reflections, homogeneous surfaces and over-exposure.A central advantage of this two-stage strategy is that we will have a rudimentary self-localization and semi-dense 3D reconstruction of the scene computed onboard in a manner that does not depend on the communication with the ground station.The resulting semi-dense 3D reconstruction can for example be used for an adaptive viewpoint planning: During the flight, the system can immediately identify for which locations it has not obtained sufficient visual information yet. This information can in turn be used to adapt the flight trajectory so as to assure an optimal exploration of unknown territory.