When the traditional struct from motion (SFM) algorithm is used to realize 3D reconstruction from the perspective of

unmanned aerial vehicle (UAV), in order to reduce the mismatching of feature points and the impact of moving targets on the overall sparse point cloud, the random sample consensus (Ransac) algorithm is mainly relied on. However, these problems can lead to a decrease in the accuracy and an increase in the number of iterations of Ransac

when solving camera poses. This article conducts target detection based on a deep learning single shot multibox detector (SSD) network. Firstly, feature points within the range of dynamic target categories are removed after scaleinvariant feature transform (SIFT) extraction of feature points. Then, mismatches are removed after K-nearest neighbor (KNN) violent matching to reduce feature points within the range of invalid moving targets and mismatching

between different categories. So that when the confidence is the same, the number of iterations of Ransac when

solving camera pose is reduced, and the time of feature point violence matching and SFM algorithm calculation of

3D points are also reduced. Finally, the feasibility of the 3D reconstruction algorithm optimized by deep learning

was verified through 12 images of 2 scenes.

unmanned aerial vehicle (UAV), 3D reconstruction, target detection, single shot mutibox detector (SSD)