Research On Building Extraction From Dense Matching Point Clouds Of Oblique Images

Dense matching point clouds are widely concerned because of their capability of representing rich ground information and have been applied more and more with the development of multi-look photographic measurement technology.However,concerning the characteristics of dense matching point clouds,such as noisy data distribution,fuzzy boundary of ground objects and high data redundancy,the traditional processing techniques for laser scanning point clouds might behave poorly on dense matching data.Therefore,it is of great application value to study the processing algorithm for dense matching point clouds.This paper mainly studies the data characteristics,elevation accuracy evaluation,filtering,building extraction,etc.from the dense matching point clouds.The main tasks and innovations are as follows:1.The research background and significance of the paper are introduced.The state-of-the-art of point cloud processing algorithms and the development of oblique photogrammetry are summarized.The characteristics of dense matching point clouds and traditional laser scanning point clouds are analyzed in detail.2.This paper designs an algorithm to evaluate the elevation accuracy of dense matching point clouds with reference of laser scanning point clouds in case of no ground control points.Firstly,the roof surfaces with stable position and geometric structure are extracted by the point-based region growing segmentation.Then the plane equation is constructed and the elevation value is fitted.Finally,the mean residuals and the Root Mean Square Errors(RMSEs)of the smallest plane element are calculated.Experiment shows that the mean residuals and RMSEs can indicate the elevation deviation and noise level of the point clouds.3.Aimed at complex urban scenes,this paper proposes a filtering algorithm to improve region segmentation of depth image and merge ground points based on semantic information of the neighborhoods.Firstly,according to the characteristics of dense matching point clouds,the virtual grid size is adaptively set based on the density of planar point clouds.Secondly,the threshold for regional growing condition is estimated based on the slope filtering algorithm.Aimed at the ground points in the enclosed area,the neighboring semantic information is analyzed along the main and sub directions of the building in the experimental area.Finally,the ground points are completely segmented.Experiment shows that the algorithm shows good adaptability to the data with fuzzy bottom edges and different density data,which leads to less artificial interaction and good filtering effect.4.A building extraction algorithm from dense matching point clouds through multi-level segmenting is proposed.Concerning the connection between buildings and trees,firstly point clouds are segmented based on the color information and multi-scale normal difference of dense matching point clouds.Then tree points are removed by establishing neighborhood voting based on the distribution characteristics.Then the top point of the grid is extracted through organizing point clouds by virtual grid.Point clouds with different topographical structure are segmented by the region growing algorithm.The plane ratio feature is used to describe the largest planar feature within the segmentation cluster.Finally,over-segmentation results are merged based on a priori knowledge of building structure.The feature vector is input into support vector machine,and the characteristics of the facade points are identified based on the extraction result of the top point in the virtual grid.Building points can be extracted completely.