| With the development of UAV(Unmanned Aerial Vehicle),navigation and positioning,UAV remote sensing as an emerging remote sensing technology is playing an increasingly important role in emergency mapping and small-scale fine mapping.How to use UAV images and auxiliary information to recover the 3D scene model is one of the key tasks of UAV remote sensing.It is widely used in autonomous driving and emergency rescue,and has important research significance and application value.With the development of UAV and camera,UAV is able to capture larger-scale and higher-resolution images,which limits the calculation time and memory consumption of 3D reconstruction.For the sparse reconstruction of long sequence UAV images,the traditional incremental reconstruction has the problem of error accumulation.Serious error accumulation will cause the bundle adjustment to converge to a local minimum or fail to converge,resulting in ambiguity.For the dense reconstruction,traditional view selection methods only consider the number of matching points and do not consider the distribution of matching points.Therefore,it is easy to select an image with a low overlap rate but very rich texture as a reference image,which will result in bad reconstruction results.Aiming at the above problems,we optimize the 3D reconstruction algorithm based on the characteristics of UAV images.We propose a hierarchical sparse reconstruction algorithm based on scene graph partitioning.The scene graph is constructed using visual consistency and spatial consistency and partitioned by normalized cut.Thus,the sparse reconstruction of largescale images is divided into multiple small-scale reconstructions,which effectively reduces the processing time of bundle adjustment.In addition,relative and absolute GPS errors are added into the cost function of bundle adjustment to constrain the camera trajectory,so as to reduce the cumulative error.We also construct a multi-scale dense reconstruction network based on group-wise correlation.We use multi-scale features and group-wise correlation to reduce the number of feature channels and depth channels,thereby reducing the memory consumption of cost volume and regularization network.At the same time,we calculate the visual correlation based on the multi-scale grids to select the appropriate reference images,and optimize the depth map filtering based on the visibility estimation,so that the dense reconstruction network can better adapt to the dense reconstruction task of UAV images.We have conducted experiments on multiple UAV datasets.The proposed optimization algorithm is better than other algorithms in terms of running speed and memory consumption.Experiments on relative and absolute GPS constraints and the view selection based on multi-scale grids also achieved good results.The experimental results fully prove the effectiveness and robustness of the algorithm. |
