Research On Large-scale 3D Point Cloud Registration Technology

With the development of science and technology,it is more and more easy to obtain 3D point cloud data which can be widely used.Thanks to the active near-infrared scanning of laser scanner,the collection process of point cloud data is not easily affected by light,weather and other environmental factors,and has strong practicability.It is widely used in transportation,road planning,road maintenance,map navigation,automatic driving,digital city and basic surveying and mapping and other fields.The point cloud data obtained at different times and directions are often based on In the independent local coordinate system,even if all the positioning information is unified into one coordinate system,there will be errors between point cloud data due to positioning failure.Therefore,most point cloud data need to be registered and spliced before processing.At present,scholars at home and abroad have put forward a lot of 3D point cloud registration algorithms.However,due to the characteristics of large-scale point cloud data,such as massive,complex,irregular,and diverse scenes,it is still facing many challenges to quickly and accurately combine the point cloud data scanned at different times and different directions into a complete 3D scene.Therefore,it is of great practical significance to study the large-scale point cloud registration technology in this paper.In this paper,aiming at the registration technology of large-scale 3D point cloud,the detailed research and practice are carried out.The main research contents and innovations are as follows:1.The related technologies of point cloud registration are summarized.This paper summarizes the characteristics of large-scale point cloud data,data organization,registration technology,registration evaluation methods and indicators and other related theories.2.A point cloud registration algorithm based on line feature matching of cumulative projection is proposed.This paper presents an automatic point cloud registration algorithm based on the cumulative projection image,which takes ?,?,?,?x,?y,?z as the medium,The six parameters of are solved in two groups.Firstly,the point cloud data to be registered and the template point cloud data are accumulated and projected in the xoy plane respectively to get the accumulated projection map,and then the first group of parameters,?,?x,?y,are obtained by using the line feature matching algorithm to register the image,Then,the same line feature matching algorithm is used to match the two images to get the second group of parameters ?,?,and ?Z.finally,the combination of the two groups of parameters is applied to the point cloud data to be registered to get the matching results.Experimental results show that the algorithm has high registration accuracy and fast processing speed in vehicle point cloud data registration.3.On the basis of line feature matching algorithm of cumulative projection,in order to solve the problem of poor feature and easy matching failure of cumulative projection in the xoy or YOZ plane,an improved point cloud automatic registration algorithm based on cumulative projection is proposed in this paper.The algorithm firstly projects the three-dimensional point cloud to the xoy plane,and then uses the SIFT algorithm to match the cumulative projection to get the matching point pair and root According to the relationship between matching point pair and 3D point cloud data,3D point cloud is obtained.Finally,the registration result is obtained by matching 3D point cloud with ICP algorithm.Experiments show that the algorithm can get the registration results quickly and accurately on the basis of high registration success rate,and has good results in vehicle and fixed station cloud data registration.4.A feature learning method of 3D point cloud based on siamese convolution neural network is proposed.On the basis of the above algorithm,in order to further improve the registration success rate and the applicability of the algorithm,this paper proposes a deep learning model based on the 3D convolution neural network.The model learns the corresponding 3D point clouds from the matching points of the cumulative projection map to obtain the 3D feature descriptors.The feature extracted by the feature descriptors and the RANSAC algorithm can effectively match the point clouds.In order to accumulate training data for our model,we project three-dimensional point cloud data of vehicle to two-dimensional plane to get the accumulated projection map,use the matching points of the projection map to the corresponding three-dimensional point cloud as the training data,and use the twin convolution neural network to learn the three-dimensional point cloud to get the three-dimensional feature descriptor.Experiments show that the descriptor can not only match the point cloud data,but also can be extended to the fixed site cloud data.Compared with other point cloud registration methods,the algorithm has a significant improvement in processing speed and registration accuracy.5.A point cloud registration algorithm based on terrain classification is proposed.According to the current point cloud classification accuracy of large-scale complex scenes is getting higher and higher,the top 5 Classification Algorithms in the open point cloud data set of typical urban complex scenes can achieve more than 90% classification accuracy for major fixed objects.We propose an algorithm based on the registration of objects after classification.This algorithm first uses kpconv point cloud classification algorithm to treat the registration point cloud and template point cloud After classification,ICP algorithm is used to register the extracted point cloud data of rod,plant and building to get the registration results.Experiments show that the algorithm can get better results in different objects registration in mobile point cloud data registration,and the fixed site cloud data can also get better results by using the extracted plants.Compared with other point cloud registration methods,this method can get effective registration results in different types of complex scene point cloud data,making the point cloud registration algorithm The applicability has been greatly improved.