Research And Implementation Of Three-dimensional Reconstruction From Aerial Images

This dissertation mainly focuses on monocular-based image sequences and binocular-based depth image three-dimensional reconstruction. The former kind of reconstruction is based on image sequences acquired from our laboratory’s Unmanned Aerial Vehicle(UAV) which equipped with an aerial camera, and to study rapid and efficient three-dimensional reconstruction of the unknown terrain or environment buildings with these images. The purpose of binocular-based depth image three-dimensional reconstruction is solving the problem of lacking the navigation information when the UAV flying at low-altitude. It can provide reconstructed environmental information and obstacle information for the assist of the UAV’s navigation. The contents of this dissertation are as follows:(1) Hardware platform and simple obstacle environment build. Binocular-based three-dimensional reconstruction algorithms will ultimately run on a DSP development board Installed in a UAV, so we first build an Embedded Development Platform and Binocular Platform. In addition, in order to verify three-dimensional reconstruction and distance measurement algorithm, this dissertation build a simple environment with obstacle to simulates real UAV flight environment, and create conditions for the verification of algorithms.(2) Camera calibration and compile of feature matching algorithms. The camera calibration operation must be carried out before the use of binocular camera, because the binocular camera’s calibration accuracy directly affects the accuracy of the match; fisheye camera distortion correction is also necessary for monocular aerial image sequence. For binocular reconstruction, this dissertation first studies how to use semi-global matching algorithm to obtain the depth image, then use triangulation principle to reconstruct the three-dimensional space point cloud, and finally shows the three-dimensional point cloud result with the mixed use OpenCV and OpenGL programming. For monocular image sequence reconstruction, this dissertation studied the traditional SIFT feature extraction algorithm and Harris corner extraction algorithm, and then use structure-from-motion and bundle adjustment to reconstructed point cloud, and finally use PMVS algorithm recovery and construct three-dimensional scenes.(3) Binocular-based three-dimensional reconstruction and obstacle avoidance system implementation. This dissertation integrated the binocular match, depth map computation, triangulation, spatial point recovery, obstacle distance calculation algorithm together to complete a three-dimensional reconstruction and obstacle avoidance system, which can provide three-dimensional environmental information and real-time distance information for UAVs.(4) Monocular image sequence reconstruction system implementation. Monocular image-based reconstruction process is cumbersome and complex, this dissertation integrated image input, feature point detection and matching, sparse point cloud reconstruction, bundle adjustment, dense point cloud reconstruction and the final output of the three-dimensional model and completed an aerial-image-sequence-based three-dimensional reconstruction system. The system can carry out three-dimensional reconstruction efficiently and automatically and the results can be outputted and saved.