Dense Matching For 3D Reconstruction With Multi-view Remote Sensing Image

In recent years,the rapid development of computer technology,especially the visual theory based on multi-view geometry,had a profound impact on the traditional mapping and remote sensing technology.For the multi-view remote sensing image,the traditional matching methods are based on generally the feature points.The main task of the traditional methods is not for three-dimensional reconstruction,but for the orientation and position of each carema through some constraints generated by a few sparse feature coreference points.And the point cloud generated by traditional methods is just sparse point cloud.However,the main purpose of dense matching is almostfor three-dimensional reconstruction and visualization tasks by every pixel.Aiming at the princeple of remote sensing image imaging and the characteristics of dense matching technology and multi-view remote sensing image,two kind images,frame array and line array,are studied in this dissertation.Two novel dense matching algorithms are proposed,and three-dimensional reconstruction experiments are carried out in public data sets,multi-view images and satellite remote sensing images.The experimental results show that the algorithm proposed in this dissertation is very good at the three-dimensional reconstruction by multi-view remote sensing images.The main innovative work is as follows:1.A dense matching algorithm based on Opponent-SIFT and Segments-Tree filtering algorithm is studied.In this dissertation,the Opponent-SIFT operator has the insensitive of the illumination of the light source and the high efficience to generate the dense descrpition.The Opponent-SIFT operator can enhance the robustness of the cost function and improve the precision of the dense matching.In the process of cost aggregation,the Segment-tree filtering is applied to non-local cost aggregation,and the initial disparity is extracted.In the process of disparity refinement,the mismatched points are corrected by the interpolation.Finally,disparity map based on thedense matching can be obtained.2.An approach based on Patch-Match algorithm for dense matching for line array image is proposed.Beacause the linear array image lacks the epipolar constraint and the matching search range is too large,the Patch-Match search algorithm is applied.This method has the multi-resolution hierarchical search strategy for dense matching.Model propagation and random searching makes the approach very efficient and adaptive.3.A three-dimensional reconstruction of multi-view frame image based on dense point is proposed.After selecting the initial reference model,the framework incrementally connects the rest of the models and cameras with the reference model by overlapping the regional model points.Finally the three-dimensional reconstruction model of the whole region could be accquired.4.A three-dimensional reconstruction processing framework for multi-view line array remote sensing image is proposed.Based on the physical model of linear array image,the dense coreference points are initial inputs,and then the spatial intersection and triangulation are carried on this framework.Finally the three-dimensional reconstruction of the terrain is formed.The above research methods and technical frameworks are experimentally verified and evaluated for accuracy and discussion.