Research On Automatic Point Cloud Registration Method Based On Key Point Matching

In recent years,the rapid development of artificial intelligence and Internet of things technology has driven great changes in various fields,such as intelligent manufacturing in industry,smart medical treatment in medicine,smart city in urban planning and smart home in people's life.Computer vision technology plays an important role in every intelligent system.Three-dimensional point cloud registration technology is one of the important technologies in computer vision field,which is mainly applied to the automatic three-dimensional reconstruction of objects or scenes.Under the background of increasing demand for intelligent applications,it is of great significance to conduct in-depth research on three-dimensional point cloud registration technology.Although domestic and foreign scholars have put forward many different technical methods in this field,they mainly rely on the classical iterative closest point(ICP)method in most practical applications.This method searches for the nearest corresponding point on the target point cloud to obtain the corresponding transformation parameters,and iterates the process continuously to complete the point cloud registration.However,ICP algorithm has a strong dependence on the initial position between the point-in-time clouds.When the initial position of two-point cloud data is poor,the convergence speed of ICP algorithm is slow and easy to converge to the local optimal,which is also an important reason for the decrease of registration efficiency and accuracy.This paper is based on the ICP registration method to carry out in-depth research and put forward some improved strategies.The main research contents of this paper are as follows:Firstly,this paper proposes a method based on voxel down-sampling and key point extraction.In order to solve the problem of the ICP algorithm's dependence on the initial position,the initial registration of the original point cloud data was carried out before the ICP algorithm.In order to improve the efficiency and accuracy of the initial registration of high point cloud,at the early stage of the initial registration of point cloud,the technology of thedown-sampling of point cloud and the extraction of key points was studied.This method can quickly estimate the appropriate size of point cloud voxel through the sampling rate under the unique parameter,and then the point cloud can be sampled under the voxel.In the extraction of key points,this paper extracts the feature point set of point cloud according to the Angle between the normal vectors of different neighborhood radii of points.The extraction of key points is an important step in the initial registration of point cloud.Secondly,this paper proposes an automatic registration algorithm based on error key point pair elimination.On the basis of extracting the key points,the initial registration and accurate registration of the point cloud are completed.In the initial registration stage of the point cloud,fast point feature histogram(FPFH)descriptors are used to describe and match the key point pairs.Because the FPFH descriptors cannot describe all the key points completely and accurately,there are many wrong matching point pairs.The method is to eliminate the error points of the matching key pairs according to the property of the vector clamping Angle.Therefore,the initial registration of the point cloud is completed by using the random sampling consistency algorithm(RANSAC)for the key points after the error matching point pairs are eliminated.In the precise registration stage of the point cloud,ICP algorithm of kd-tree structure is directly used to complete the accurate registration of the point cloud after the initial registration.Finally,the improved method proposed in this paper is simulated with different point cloud data.The simulation results show that the automatic registration method is feasible and effective.