Research On Point Cloud Normal Estimation

Point cloud data is a three-dimensional coordinate information obtained by sampling a real object surface of in real world by a 3D scanner.With the application of a 3D laser scanner,more and more point cloud data can be obtained,due to its flexible acquisition method,simple model data structures,point clouds become the representation of commonly used geometric models.Among them,point cloud normal vector is the most important geometric attribute of point cloud data.Accurate normal vector estimation can be widely applied to surface reconstruction,point cloud rendering and point cloud segmentation.However,when the scanner acquires data,it is affected by the physical limitations of the device itself and the external factors of the scanned object.The obtained point cloud data is often accompanied by noise,anisotropy sampling,data missing,etc.,which can be efficient and robust.Point cloud normal estimation,which can maintain sharp features well,needs to be addressed.Many normal estimation algorithms perform normal estimation at feature points.The neighborhood of the normal structure often contains two or more smooth surfaces.The points on different surfaces participate in the normal estimation of the current point.Thus cause the big error.A normal estimation algorithm based on neighborhood segmentation adopts the method of anisotropy neighborhood on feature points.The neighborhood center point we select is at a certain distance from the current point,so the constructed neighborhood can truly reflect the neighborhood information of current point,which can maintain the sharp features well.Two specific neighborhood segmentation techniques are designed: edge point neighborhood segmentation and corner point neighborhood segmentation.An improved point-to-point consistency voting method is proposed.A new cross entropy weight is added to the previous voting function.The proposed new voting function balances the error from the point to the fitting plane,consistency of the initial normal between the pairs and consistency of the fitting error between pairs.the estimated normal performed well,which can maintain sharp features,robust to noise and overcome anisotropic sampling.