Research On Object's 3D Point Cloud Registration Method

With the continuous development of machine vision related technologies,the research on images has already risen from two-dimensional to three-dimensional.This is based on the transformation that 2D images contain less information than 3D images.Due to the opacity of the object,all the point clouds of the object cannot be collected in a single direction.Multiangle shooting is required to get the complete information of the point cloud.To stitch these data into a complete object model,point cloud registration technology is required.The existing point cloud registration algorithms have certain limitations.The registration speed with high registration accuracy is slow,and the registration accuracy with fast registration speed is low.In order to balance this problem,the rough matching of the point cloud registration algorithm The quasi and fine registration are improved respectively.The key points and their descriptors are calculated based on the conformal geometric algebra to reduce the computational time for finding the initial pose in the coarse registration.The transformation matrix can be calculated at a faster speed while still being able to Improve coarse registration accuracy.The optimization of the traditional ICP algorithm based on the iterative method of weight selection is insufficient in resisting gross errors.Although a certain registration efficiency is sacrificed,the accuracy of the registration result is very high.The main work of this paper is as follows:(1)Research the 3D reconstruction technology,analyze the role of the point cloud registration algorithm in this technology,and give directions for improving the registration algorithm.The RGB-D image of the model is collected with a depth camera,and a CAD model is generated in the computer by mathematical methods based on the collected model.Imitate the visual robot in the industry to recognize a specific object.(2)In-depth research and analysis of the coarse registration steps of the point cloud registration algorithm.Conformal geometric algebra is used as the descriptor extraction algorithm,and curvature is used as the selection criterion of feature points.Then,the initial pose between the point clouds to be registered is searched to provide input conditions for the fine registration algorithm.And compare the computational complexity and registration accuracy between this method and the feature descriptor algorithm.The experimental results show that the algorithm in this paper can reduce computational time while ensuring high accuracy.(3)In-depth research and analysis of the precise registration steps of the point cloud registration algorithm.In the fine registration process,because the ICP algorithm cannot resist gross errors,and based on the improvement of the least square method,it is proposed to add a set of weighting iterations in each ICP iteration process to determine the weight of the residual.This way of determining the weight is that the residual is large and the weight is small;the residual is small and the weight is large,which meets the requirements.And add the conventional gross error and edge gross error to the point cloud to carry out the experiment.The results show that the improved algorithm can maintain high registration accuracy regardless of whether it is dealing with conventional gross errors or edge gross errors.