Research On 3D Point Cloud Registration Algorithm Based On Image Features And ICP

3D laser scanning technology can intensively acquire a large number of data points of the target object,and can quickly reconstruct the 3D model of the measured target and various map data such as line,area,body,etc.The technology from single point measurement breakthrough to surface measurement,It is widely used in many fields such as heritage protection,architecture,planning,disaster assessment,ship design,digital city,military analysis,reverse engineering,and unmanned driving.3D point cloud data registration is an important part of the application of this technology,and also a key step of point cloud data processing.At present,with regard to the algorithms involved in 3D point cloud registration,previous researches have not yet achieved satisfactory results in balancing registration accuracy and registration speed.To calculate point cloud feature information and process a large amount of point cloud data,the current algorithm still has the problem of long registration time.When the initial pose of the point cloud is not good or there is a wrong corresponding point pair,there is also the problem that the registration accuracy is not high and even falls into the local optimal.To deal with these problems,from the perspective of improving the accuracy and speed of point cloud registration,relevant research on point cloud registration is carried out.The main contents include as follows based on the idea of “from coarse to fine”.First,aiming at the problem of conversion of 3D point cloud data to 2D image,the thesis realized the conversion of 3D point cloud to 2D bearing angle image based on the bearing angle information of 3D laser data point.The conversion principle include: first of all,establish the image plane coordinate system.Then,map the three-dimensional laser point to the image plane.finally,map the bearing angle value of the laser point to the pixel value of the image.Second,aiming at the bearing angle image feature matching problem of the 3D point cloud,an bearing angle image feature matching algorithm based on ORB and GMS was constructed.The bearing angle image of 3D point cloud was studied.And the bearing angle image feature matching is implemented through ORB and GMS.The basic ideas include:first of all,according to the above principle of conversion from 3D point cloud to 2D image,convert 3D point cloud into 2D bearing angle image.Then,a large number of 2D image features was extracted through the ORB algorithm.And the rough matching of bearing angle image features was realized through the brute force matching method based on the hamming distance-based.Finally,the GMS algorithm is used to complete the secondary matching of bearing angle image features to obtain 2D matching point pairs.Third,aiming at the issue that it consume a lot of registration time for a large amount of data and the calculation of point cloud feature information.And the poor initial position of the point cloud reduce the registration accuracy.A point cloud rough registration algorithm,based on bearing angle image features,is established.This algorithm comprehensively considers the characteristics of 3D point cloud feature matching and 2D image feature matching.The basic ideas include: First of all,according to the one-to-one correspondence between bearing angle image pixel points and 3D point cloud data points,the corresponding point pair set of the 3D point cloud are extracted through the bearing angle image matching point pair set.Then,based on this corresponding point pair set,the point cloud rough registration is performed by using the singular value decomposition method.Fourth,aiming at the problem that the ICP algorithm consume a lot of registration time in large data,and the accuracy of ICP algorithm was affected for the mismatched pairs.A point cloud fine registration algorithm,based on improved ICP,is established.According to the four aspects: point cloud reduction,accelerated search for corresponding points,elimination of mismatched pairs and solution of transformation matrix parameters,the his improved ICP algorithm is obtained.The basic ideas include: first of all,the voxel grid is used to reduce the original points.This measure can reduce the amount of data involved in the corresponding point search,thereby improving the registration speed.Then,the KD tree is used to organize simplified point cloud data,and the corresponding points pairs are searched based on KD tree,thereby speeding up the search of the corresponding points.Next,the RANSAC algorithm is used to eliminate the mismatched pairs to obtain a purified corresponding point pair set.Finally,for the purified corresponding point pair set,the dual quaternion method is used to compute the transformation matrix parameters.And the transformation matrix is used to complete the point cloud fine registration.Finally,according to many point cloud data sets,the above theory are studied andanalyzed.In terms of the conversion of a 3D point cloud to a 2D image,a conversion scheme based on information such as the bearing angle of the 3D laser point is effective.In terms of bearing angle image feature matching,compared with SIFT+GMS,BRISK+GMS,AKAZE+GMS,and ORB+RANSAC algorithms,the algorithm has more advantages such as the number of feature points,detection time of the same number of feature points,the number and time of secondary matching pairs,and the corresponding point cloud registration accuracy.Therefore,the algorithm has better performance in feature extraction and matching.In terms of rough registration,compared with FPFH and 3DSC algorithms,the CPPDE and registration time of rough registration algorithm are smaller,which has better registration effect.In terms of fine registration,the improved strategy,based on voxel grid,KD tree,RANSAC and dual quaternion,reduces pairs distance error and time of registration.And the fine registration algorithm can effectively improve registration accuracy and speed.The experimental results show that the CPPDE and registration time of our fine registration algorithm is better than ICP and NDT algorithms,which has better registration effect.