基于即时稠密三维重构的无人机视觉定位

传统景象匹配定位方法在用于低空无人机定位时，易因低空航拍图像视场小，且与卫星图像（带有地理信息）的拍摄角度差异大而失败.本文提出了一种基于即时稠密三维重构的无人机视觉定位方法，通过将稠密三维点云与卫星图像匹配以实现无人机定位.首先根据图像序列快速估计摄像机位姿，而后使用多深度图协同去噪与优化算法生成稠密三维点云，随后通过变换观察视角由稠密三维点云生成与卫星图像拍摄视角相近的虚拟视图，最后将虚拟视图与卫星图像匹配并得到无人机的地理坐标.由于稠密三维点云包含多张图像的信息，覆盖面积大，且可变化观察视角，因此能够有效克服上述两个问题.实验证明了本文方法的有效性.

Instant Dense 3D Reconstruction-Based UAV Vision Localization

The corresponding matched images are hard to obtain by the convetional scene matching localization method applied in UAV (Unmanned Aerial Vehicle) at a low altitude because of the small viewing coverage and the big capturing angle difference comparing with the satellite images.We propose a localization method based on instant dense 3D reconstruction.Firstly,we use a fast SLAM (Simultaneous Localization And Mapping) method to retrieve camera poses of image sequence captured by UAV.Secondly,cooperative denoising and optimization algorithm across multiple key frames is applied to obtain dense depth map and dense point cloud.Thirdly,a virtual view,the angle of which is similar to that of satellite,is generated by iterative optimizing method.Finally,we estimate the position of UAV by correspondence between the satellite map and the previous generated virtual view.Since the dense 3D point cloud integrates the information of multiple aerial images with small field of view and the viewing angles of some generated virtual views are close to those of satellite images,the proposed method provides a higher success rate and accuracy for localization.Experimental results illustrate the effectiveness and applicability of the proposed framework.