Research On Semi-Global Dense Matching Algorithm For Aerial Images Using Edge Information

The advantages of digital photogrammetry have made it an important technical means in urban 3D modeling.The dense matching of aerial stereo images is to find the correspondence points in the stereo pairs,and then use triangulation to obtain the threedimensional space coordinates of the corresponding ground points.Dense matching of aerial images can reconstruct dense 3D point clouds with texture,color,and position information.Dense point clouds have become an important data source for 3D modeling.However,the coverage of aerial images is large,the scene of aerial images is complex,and there are problems such as image noise,lighting changes,textureless and repeated texture,occlusion and depth discontinuity,which bring great challenges to dense matching.Semi-Global Matching(SGM)has been widely used in photogrammetry due to its dual advantages of high precision and high efficiency.However,there are still some problems that matching cost cannot describe the real difference between correspondence points and the fixed penalty parameters cannot satisfy the matching effect of both flat regions and depth discontinuities simultaneously.Due to the smoothness assumption in the matching model,there is a problem that mismatch rate in depth discontinuities such as the edge of houses is high,which leads to loss of the characteristics of houses,the blurring of edge,and the extension of foreground disparity to the ground.Although many researchers have put forward many methods,none of them can solve the problem completely.In response to the above problems,this paper studies the following under the framework of Semi-Global Matching algorithm:(1)By analyzing the performance of different matching costs in aerial images,performance evaluation indexes include the accuracy of disparity maps and the visual effects of point clouds,it is found that the matching cost determines the accuracy of SGM.Experiments show that,in SGM for aerial images,compared with other traditional matching costs,Census appears overall to be the most robust and most accurate cost,and the matching cost based on the gradient performs well at depth discontinuities.(2)In response to a series of problems existing in the traditional Census transform,this paper uses the gray mean of neighborhood pixels in the Census transform window to replace the gray value of the central pixel,so as to overcome the matching error caused by sudden change of gray value of the central pixel affected by noise,and then introduces a gradient operator to further increase the comprehensive performance of the Census cost.Experiments show that the matching cost of the fusion of improved Census transform and gradient is more robust than the traditional Census cost,and edges are better maintained at depth discontinuities.(3)In order to deal with the problem that the fixed parameters in SGM are difficult to adapt to different terrain conditions,this paper uses the prior knowledge that depth discontinuities often correspond to edges of objects,and uses edge detection method based on the point mutual information for obtaining edge information of images,and then adaptively selects penalty parameters according to the edge attributes of each pixel: imposes a small penalty on edge pixels,impose large penalty on non-edge pixels.Finally,these parameters are used in SGM.Experiments show that the proposed algorithm improves the matching accuracy of flat areas and depth discontinuities.The generated point clouds not only are more complete on flat areas such as ground and roof,but also have more complete geometric features of houses and sharper edges.