Research On Registration Algorithm Of The Point Cloud Data Based On Structural Features

In reality,shape matching and image registration are fundamental problems in image analysis,computer vision and pattern recognition.Because point representations are common and easy to extract,they are often emulated as a point cloud registration problem.Meanwhile,the point cloud registration is the core technology of three-dimensional reconstruction,and it is the hotspot and focus in computer vision field such as simulation design,virtual reality and so on.Moreover,with the development of modern sustainable technology,the computer hardware technology has been improved steadily,which also makes the collection and processing of digital model have enough hardware support and provides sufficient technical support for the research of the point cloud registration.The aim of the point cloud registration is to determine the transformation relationship between two point clouds,in other words,to map one point set to another by a series of transformations,and to obtain a one by one correspondence between the elements of two point clouds.According to the different transformation relations between the point sets,the point cloud registration can be roughly divided into two categories: rigid point cloud registration and non-rigid point cloud registration.The rigid point cloud registration contains a small number of transformation parameters,such as translation,rotation and scaling,therefore,rigid registration is relatively easy and has been studied relatively mature.But the non-rigid transformation is a more complex transformation,it not only contains the transformation parameters in the rigid registration,but also includes such as telescopic,affine,projection,polynomial and some of the more complex transformations.Since the real non-rigid transformation is often unknown,the non-rigid registration is relatively difficult.In real life,the examples of non-rigid registration are widespread,so it is very important to study the non-rigid registration.First of all,this paper elaborated the significance of point Cloud registration method research as well as the domestic and foreign research present situation,the paper also introduces the important principles of point cloud registration including probability density model,geometric transformation model,probability density function estimation algorithm,especially the Gaussian mixture model,asymmetric Gaussian mixture model,Some common geometric transformation models and EM algorithm.Secondly,focusing on the target of the non-rigid registration,the main works of this paper are as follows:(1)This paper presents a point cloud structure feature descriptor combining distance and angle information.We introduce the structural features of point cloud into the registration process by this descriptor,and propose a method of non-rigid point cloud registration based on Gaussian mixture model and local structure feature.In this method,a point set is used as the center of mass of Gaussian mixture model,another point set as the data point of Gaussian mixture model is used to fit the Gaussian mixture model.The point cloud registration problem is transformed into a probability solution process,and a better registration result is obtained.(2)This paper proposes a GLSC descriptor of combining local and global structure based on the shape context descriptor.The descriptor has a stronger descriptive effect on the structural characteristics of the point cloud.Meanwhile,considering that although the single Gaussian model and the Gaussian mixture model are often used in many research fields of machine vision and pattern recognition,they can't fit all the distributions in the reality.The point cloud model collected in reality has asymmetric distribution.In this paper,the asymmetric Gaussian model is introduced,and a large number of experiments have proved that the asymmetric Gaussian model can capture the spatial asymmetry distribution.By analogy Gaussian mixture model,this paper constructs an asymmetric Gaussian mixture model to fit the point cloud model,and presents a method of non-rigid point cloud registration based on asymmetric Gaussian mixture model and GLSC descriptor.