Research On The Method Of 3d Surface Reconstruction Based On Multi-view Dense Matching Point Clouds

Three-dimensional reconstruction based on multi-view images is one of the research hotspots in photogrammetry and computer vision in recent years.Oblique photographic technology developed for the three-dimensional reconstruction of the urban area,can generate fine three-dimensional model of city based on large-scale,high-resolution and multi-angle aerial imagery.However,the shortness of baseline between cameras in the imaging system,sever projective distortions in oblique images.serious occultation of sight and other issues,which can directly affects the completeness and precision of 3D model reconstruction.Therefore,we focus on some critical problem of 3D reconstruction from oblique images in the paper,and propose our idea base on three aspects,such as multi-view aerial images dense matching point clouds acquisition,subsequent point clouds processing and geometric model reconstruction.The content of this paper is showed as follow:(1)First of all,the technology and principle of 3D reconstruction based on image are collated.On basis of stereo correspondence principle,we compare point clouds in terms of density,accuracy,completeness,which are obtained through different stereo matching algorithms.Finally,the improved SGM algorithm is chosen as the dense matching algorithm for point clouds generation from multi-view aerial images.(2)We propose a suitable strategy to deal with multi-view images dense matching point clouds,and improve the overall quality of point cloud.The strategy include the elimination of redundant point clouds,point clouds smoothing and resampling operation.By analyzing the causes and distribution characteristics of point clouds noise,noise detection is carried out by using two methods of back projection error and distance statistical analysis.Considering the idea of image median filtering,we can remove the redundancy of point clouds in the area which can be seen in more than two images.The improved moving least squares method is used to smoothing the local flat area with the uneven point clouds.As the weak or none texture regions in images are difficult to achieve exactly matching,and the elimination of redundant points will inevitably produce holes,we propose different criterions for grid interpolation to achieve uniform distribute result during digital surface model generation.(3)In comparation of mainstream algorithms of triangulation for automatic generation of digital surface model,poisson surface reconstruction is taken as a suitable one.In order to achieve high quality poisson surface reconstruction,under the premise of point clouds preprocessing and a reasonable normal filed distribution after normal consistency,the surface reconstruction problem is transformed into poisson problem,and can be solved by implicit function in mathematics.Further more,the limitations of the poisson reconstruction algorithm can be approved in respect of anti-noise property and data redundancy.Finally,we do some experiments for digital surface model automatically generation from oblique images and UAV images respectively.As results,the proposed method can improve the completeness and precision of the digital surface model to a certain extent,and satisfy the application requirements.