Research On The 3D Point Cloud Registration Based On Depth Camera

With the rapid development of emerging technologies such as driverless,AR / VR,cultural relics restoration,and face recognition,3D modeling has become a research hotspot for researchers.3D point cloud data has become a common processing object in 3D modeling because of its flexibility and ability to build complex and detailed models.As a new 3D representation form,3D point cloud data contains rich geometric topological information,but it also has many defects.In order to solve the existing defects,the 3D point cloud processing technology has been developed.The 3D point cloud processing technology mainly includes point cloud denoising,point cloud repair,point cloud feature extraction,point cloud segmentation,point cloud registration,point cloud surface reconstruction and other processes.This subject mainly studies comparison and analysis of different point cloud registration algorithms in coarse registration and complete registration.The main research works of this topic are included as follows:(1)The camera calibration of the Kinect v2 depth camera is completed,the internal parameters,the external parameters and the distortion parameters of the camera are obtained.The average pixel error is used to evaluate the obtained calibration results,and the evaluation results meet the error requirement.(2)The comparison and analysis of different point cloud registration algorithms in coarse registration are completed,six different point cloud registration algorithms are used to perform coarse registration experiments on rabbit point cloud models with low coverage,dragon point cloud models with large noise data and buddha point cloud models with large amount of data which have good experimental performance.By comparing the time which is spent in the registration process,the registration efficiency of different point cloud registration algorithms can be obtained.By comparing the point cloud model after registration with the target point cloud's rotation angle error and translation distance error on the X,Y,and Z axes,the following conclusions can be obtained: For rabbit point cloud models with low coverage,the point cloud registration accuracy of ICP is significantly lower than that of other registration algorithms;For dragon point cloud models with large noise data,the point cloud registration effect of NDT is poor;For buddha point cloud models with large amount of data,the point cloud registration accuracy of ICP is better,while the point cloud registration accuracy of NDT still performs poor.Finally,the rotation angle error and translation distance error are compared.A large difference can be obtained with respect to the rotation angle error,and the translation distance error is almost the same.(3)The comparison and analysis of different point cloud registration algorithms in complete registration are completed,ten different point cloud registration algorithms are used to perform complete registration experiments on rabbit point cloud models with low coverage.By comparing the time which is spent in the registration process,the registration efficiency of different point cloud registration algorithms can be obtained.By comparing the point cloud model after registration with the target point cloud's rotation angle error and translation distance error on the X,Y,and Z axes,the following conclusions can be obtained: In the complete registration with ICP as the fine registration,the algorithm based on local features as the coarse registration has the highest accuracy;In the complete registration with NDT as the fine registration,the algorithm based on four-point consensus set as the coarse registration has the highest accuracy.The rotation angle errors of complete registration with ICP and NDT as fine registration are compared,it can be obtained that: In the case of ICP as the fine registration,the registration accuracy of the complete registration with 3DSC,PFH,FPFH and 4PCS as the coarse registration has little difference in the rotation angle error in the X,Y,and Z axis directions.When NDT as the fine registration,the rotation angle error of each point cloud registration algorithm has a large difference;In the case of the same kind of coarse registration,the overall registration accuracy with ICP as fine registration is higher than that with NDT as fine registration.Based on the above researches,this paper can get the registration characteristics of different point cloud registration algorithms in different scenes.When choosing a coarse registration that provides an initial transformation matrix to the point cloud model,the point cloud registration algorithm based on local features can be preferred,because this kind of algorithm has high registration accuracy in coarse registration,ICP and NDT are not suitable for coarse registration;When choosing a complete registration with higher registration accuracy for the point cloud model,a complete registration when ICP as the fine registration can be preferred,because the accuracy and stability of this type of algorithm are better than that of NDT as the fine registration;Because of the visual display of the rotation angle error and translation distance error of each point cloud registration algorithm on the X,Y,and Z axes,it can provide good direction and idea for the optimization and improvement of the point cloud registration algorithm in the future.