Research On Point Cloud Normal Estimation And Alignment

With the development of 3D scanner,more and more 3D data are available.By themeantime,the need and reliance of 3D scanner technology is growing stronger in the industry.Thus,the process of point cloud data has been a significant topic for the domain of computer graphics.However,the point cloud data obsessed with 3D scanner is hard to use directly.First of all,there might be noise generated during the scanning procedure,also the point density of cloud might be not uniform across the surface of object.Constrained by the camera's angle,each frame of scanned data might only cover part of the object's surface.The point cloud is a scattered kind of geometrical data.The connection between points has be lost,hence the geometrical information is incomplete on the point cloud.To retrieve the full information,generally we reconstruct point cloud into mesh.There are a fewer preprocess procedure need to be done before reconstruction.First,we need to merge the partial point cloud data of each frame into a full model,this procedure is called point cloud alignment.Second,many surface reconstruction methods require subtle and accurate normal.Consequently,the efficient normal estimation technology is of great importance.In this paper,we mainly focus on the technology point cloud estimation and alignment.We firstly proposed a normal estimation algorithm using shifted anisotropy neighborhood and proved is efficiency using experiments.Then we proposed a PCV-MN method to find multi-normal for points around features,we define that point lies on the intersection of surfaces has multi normal.After the PCV-MN method is introduced,we followed with an extensive point cloud estimation benchmark,which compared classical,state of art,and our algorithm on a wide covering data base.The benchmark demonstrated the advantage of our algorithm.At last we introduce some work on point cloud alignment,including making alignment from correspondences of 2D texture,and point cloud alignment using deep learning based feature landmark.