Research On Fast Point Cloud Registration Algorithm Combining Global And Local Features

3D reconstruction based on RGB-D image is currently a hot research issue.The accuracy of the reconstructed object obtained by this method depends mainly on the registration accuracy of the point cloud obtained from each perspective.The reconstruction method based on RGB-D image restores the three-dimensional structure of the target according to the motion information of the depth device.In this method,the depth equipment is used to directly acquire the point cloud data of each view point during the movement,and then analyze the relationship between the point clouds of two adjacent viewpoints to estimate the transformation matrix between the adjacent positions d uring the depth device movement.Then the point cloud of each perspective is transformed into the same reference coordinate system according to the above transformation to get the complete point cloud.Finally,Poisson surface reconstruction method is used to process the complete point cloud to get the final reconstruction result.In 3D reconstruction based on RGB-D image,a large number of point cloud registration algorithms restrict the registration of point cloud by analyzing texture information and depth information,which results in unsatisfactory result of point cloud registration with missing texture information.However,the accuracy of point cloud registration by analyzing the structural relationship between point clouds will decrease sharply with the absence of point clouds.Therefore,a fast point cloud registration method is proposed in this paper,which not only is not affected by the target texture information,but also can complete the point cloud with high registration accuracy when the degree of missing point cloud is as high as forty percent.Firstly,this paper proposes a global nearest and farthest point registration algorithm based on center of gravity of the point cloud.The algorithm firstly calculates the center of gravity of the original input point cloud,and then uses the down-sampling to thin out the original input point cloud to obtain the corresponding sparse point cloud.Then according to the characteristics of partial overlapping of point cloud structure of two adjacent viewpoints,the nearest6)₁ points are found in the corresponding sparse point cloud with reference to the center of gravity of the point cloud,and then the6)₁ points are used as the reference points in the sparse to find the nearest and farthest6)₂ points in the point cloud.The angle between the nearest point,the reference point and the farthest point is regarded as the initial seed matching feature of the adjacent point cloud.Then the other matching points are searched from the obtained seed matching neighborhoods,and then the transformation matrix between adjacent depth devices is computed according to the matching point pairs to obtain the final registration result.Then,in order to ensure that the registration algorithm can still be effective in the case of a high degree of missing point clouds,this paper proposes a nearest and farthest point registration algorithm based on local reference points.The algorithm firstly down-samples the original point cloud to obtain a sparse point cloud,and then each point in the sparse point cloud is used as a reference point to find its neighborhood in the or igin cloud,and its nearest and farthest6)points in the neighborhood are found.Then the angle between the nearest point,the reference point and the farthest point is regarded as the seed matching feature of the adjacent point cloud,and then other matching point pairs are obtained from the neighborhood of the seed matching,and then the transformation matrix between adjacent depth devices is computed according to the matching point pairs to obtain the final registration result.In summary,this paper proposes two methods of point cloud registration,the former based on the global features and the latter based on the local features.When the number of missing points is small,the global matching algorithm is used to complete the registration quickly.On the contrary,when the number of missing points is large,the local registration algorithm is used to make up for the lack of global registration.The registration algorithm having been tested on several datasets,can quickly and exactly complete the registration.