| The demand for the surveying and geographic information industry for city-level real 3D city construction has resulted in an increasingly widespread use of unmanned aerial vehicle(UAV)images for 3D reconstruction.Currently,the theory and technology of 3D reconstruction based on point feature have been developed more maturely,while the research and application of 3D reconstruction based on line feature are relatively limited.Line features are widely distributed in human-made structural scenes such as cities.They not only express the structural information of the scene,but also can still be useful in some complex scenes that are susceptible to occlusion.Therefore,it is of great value to study the theory and technology of 3D reconstruction based on line features and incorporate line features into the theoretical system of traditional point features 3D reconstruction.Accurate and reliable results of pose estimation are important guarantees for the realistic expression of real-scene 3D models.Therefore,image pose estimation becomes a key part of the 3D reconstruction data processing.This thesis focuses on the pose estimation method of UAV images in city scenes.From the perspective of point and line feature fusion processing,this thesis focuses on line feature matching of wide-baseline stereo images and camera pose estimation and optimization.The main research work is as follows:(1)Summarize the research status and problems to be solved at home and abroad.Based on the background and significance of point-line feature fusion processing,the thesis focuses on the current research status of wide baseline stereo image line feature matching and camera pose estimation,and summarizes the problems that wide baseline stereo image line feature matching is easily affected by the large viewpoint difference of image and the unreasonable weighting method of point-line feature fusion estimating pose.The thesis introduces the theories and techniques related to camera imaging model,point feature matching,line feature matching and UAV image pose optimization in detail with data acquisition and processing as the inner clues,and focuses on the key aspects of the UAV image pose estimation-the nonlinear least squares optimization method and the variance-covariance component estimation theory,including Gauss-Newton method,Levenberg Marquardt(LM)optimization method and Helmert variance component estimation.This provides the theoretical basis for the solution of the above problem.(2)Aiming at the problem that the large viewpoint difference of wide baseline stereo images affects the matching of line features,a point feature guided matching method for wide baseline stereo image line features is proposed.This method compensates the local affine deformation of the image by simulating the possible viewpoint changes of the camera.Firstly,the fundamental matrix between images was calculated based on the successfully matched ASIFT(Affine Scale Invariant Feature Transform)points.Secondly,the local affine deformation of the original image was corrected and image pyramid was constructed to generate the descriptors of LJL(Line-Junction-Line)structural elements in the multi-scale image space.Finally,based on the epipolar geometry between image pairs,the search range of junction in the LJL structural element is reduced,and evaluate the similarity between line feature descriptors using Euclidean distance.Three sets of experimental data were used to assess the matching performance of the method from three aspects: narrow baseline stereo image pairs,same-flight height wide baseline stereo image pairs,and differentflight height wide baseline stereo image pairs.The experimental results show that the proposed method not only significantly improves the correct matching rate of line feature,but also greatly enhances the matching efficiency of line feature.The proposed method can partially eliminate the influence of image viewpoint difference,compared with line feature matching of narrow baseline stereo images with small viewpoint difference.(3)Aiming at the problem of unreasonable weighting point and line feature fusion for solving camera poses,a camera pose estimation method with adaptive fusion of point and line feature was proposed.Firstly,considering the difference in accuracy of point and line feature observations,point and line features are regarded as two different types of observations,and constructs the pose estimation model by using the optimal estimation under the least squares criterion.Secondly,the weight of the two types of observations is iteratively adjusted and the camera pose is solved accordingly,based on the principle of Helmert variance component estimation.Finally,the results are converted to the world coordinate system using ground control points.The accuracy of the proposed method is evaluated by using simulated data and real data.The experimental results show that the participation of line features in the pose estimation can significantly improve the accuracy of the pose solution.Moreover,compared with weighting point-line features equally,the adaptive weighting strategy of point and line features can further improve the accuracy of camera pose solution,thus exploring a feasible way for camera pose estimation by combining multiple features.There are 46 figures,9 tables and 92 references. |
