Research On Efficient SfM Reconstruction Of Oblique UAV Images

Because of the rapid development and wide application,UAV(Unmanned Aerial Vehicle)has become a new remote sensing platform in extensive fields.Meanwhile,oblique imaging technology is capable of capturing both the footprints and facades of targets and is considered the bridge between classical and terrestrial photogrammetry.Therefore,the UAV oblique photogrammetric system could improve both of their applicabilities for strength integration.In the field of computer vision,SfM(Structure from Motion)technique can simultaneously recover poses of images and reconstruct 3D points of scenes,which is able to be used for the sparse reconstruction of UAV oblique images.However,the primal SfM technique is designed for the reconstruction of small-scale scenes or small-size images in the field of computer vision.Thus,this study designs the image pair selection algorithm based on the analysis of topological connection network to address the combinational complexity of images;the feature extraction and matching solution based the tiling strategy to cope with the difficulty resulted from oblique imaging;the hierarchical motion consistency constraint to solve the problem of geometrical verification of initial matches with high outlier ratios.Finally,the overall workflow for the sparse reconstruction of UAV oblique images is established in this study.The primary contributions of this work are listed as followings:First,the SRC-InterTest algorithm is proposed for initial image pair selection based on spatial relationship constraints,which aims to search overlapped image pairs.In this study,we proposed the SRC-InterTest algorithm to select image pairs,which exploits two spatial relationship constraints.The first one is the Spatial Distance Constraint(SDC),which is used to achieve highly efficient intersection tests;it is independent of the fixed radius for neighbor searching and can be used to avoid exhaustive tests.The second one is the Spatial Overlap Constraint(SOC),which filters image pairs with overlapped regions that are too small or too narrow,aiming to simplify the image connections as early as possible.Second,the MST-Expansion algorithm is proposed for match graph extraction based on topological connection analysis.Vase numbers of redundant and unessential match pairs are still retained because the direct adjacent principle is used for match pair selection.Thus,we proposed the MST-Expansion algorithm for match graph extraction from initial candidate match pairs.Firstly,a topological connection network(Reduced-TCN)is constructed using the initial candidate match pairs,which is represented by an undirected weighted graph;secondly,a two-stage algorithm is used simplify the connection network:a maximum spanning tree(MST)is extracted from the Reduced-TCN to obtain the simplest form of the topological connection network;then,the MST is gradually enhanced through local structure analysis.Finally,the extracted match graph can be used to guide feature matching and decrease time costs in the stage of image matching.Third,this study analyzed the POS-assisted feature extraction and matching strategy for oblique UAV images.Considering inherent characteristics,the validation of geometrical rectification and tiling strategy for oblique images needs verification.Thus,we first analyze the influence of POS-assisted geometrical rectification and tiling strategy on feature extraction and matching of oblique UAV images in detail;then,based on these two strategies,we designed four different solutions;finally,the best solution suitable for oblique UAV image is determined by using extensive comparison and analysis of feature extraction and matching experiments.Fourth,the HMCC-RANSAC algorithm is proposed for geometrical verification of initial matches with high outlier ratios.High outlier ratios can cause low-efficiency and unstable result of the classical RANSAC algorithm.Thus,this study established a stable and efficient strategy by combing the Hough voting scheme and the RANSAC algorithm.Firstly,initial candidate matches are projected onto the object space for the simplification of geometrical transformation in the image space,which results in the simple 2D-translation in the object space;then,two characteristics of the simplified model,namely direction and length,are used as constraints to detect false matches,which is implemented by using the Hough voting scheme;finally,the rigorous verification of the RANSAC algorithm is used to refine the final matches.Finally,by using the image pair selection algorithm,the strategy for feature extraction and matching,and the geometrical verification strategy for initial matches with high outlier ratios,this study established the overall workflow for the sparse reconstruction of oblique UAV images.By using four oblique datasets for comparison tests with MicMac and Agisoft PhotoScan,the experiment results demonstrate that the proposed workflow can achieve efficient reconstruction of oblique UAV images under the preservation of accuracy and completeness.