Research On Point Cloud Registration In Different Application Scenarios

With the development and popularity of depth sensors such as lidar and depth cameras,point cloud registration technology has attracted more and more attention.Point cloud registration refers to the process of converting the point clouds of multiple scenes or objects scanned by sensors to the same coordinate system by calculating the transformation matrix,so as to obtain a complete point cloud.Point cloud registration technology has been widely applied in many fields,such as 3D object pose estimation,3D reconstruction,mapping and positioning,as well as the emerging augmented reality in recent years.According to the different characteristics of the application and corresponding sensor data,point cloud registration is divided into dense point cloud registration in a small-scale environment and sparse point cloud registration in a large-scale environment,and in-depth exploration and research are carried out respectively.The specific work content is as follows:(1)For the registration of dense point clouds with a small number of outliers,a refined registration algorithm based on parameter decoupling Gaussian Mixture Model is proposed.Aiming at the problem that the feature expression ability of the gaussian mixture model is reduced due to the high coupling of the parameters in the traditional method and the solution space is limited,the deep neural network is introduced to realize the decoupling of different parameters,and the corresponding parameters are solved by extracting the global and local features respectively.The decoupling of parameters makes the learning of different parameters independent and the solving space of each parameter is larger.At the same time,the method of solving corresponding decoupled parameters with different information makes full use of the attributes of different parameters and fully mines more abundant information in the point cloud.Experiments on the Model Net40 dataset and the ICL-CULM dataset verify the fine registration performance of the algorithm and its high generalization to new data.(2)For dense point cloud registration with a large number of outliers in a small indoor scene,a robust registration algorithm based on consistency constrains coding attention network was proposed.In this algorithm,multiple geometric constraints inherent in rigid point cloud registration are introduced,and a consistency constrains coding attention network is designed accordingly to fully learn and extract potential correlation features between point cloud information and multiple geometric constraints.Then,this feature is combined with the proposed consistency neighborhood similarity matrix to solve the point pair confidence,so as to realize the extraction of the correct corresponding point pair and the elimination of the wrong corresponding point pair.Experiments on the 3DMatch dataset and real indoor environments verify the robust registration performance of the algorithm.In addition,the refined registration algorithm based on parameter decoupling Gaussian Mixture Model is combined with the robust registration algorithm based on consistency constrains coding attention network,which further improves the registration accuracy of the algorithm.The fusion algorithm is validated on the 3DMatch dataset and real indoor environment data.(3)For the registration of sparse point cloud of large-scale scene scanning,a key point matching method based on an adaptive dynamic update graph attention network is proposed.Its registration implementation is based on key points extracted from the point cloud.This method overcomes the shortcoming of traditional graph neural network which cannot gather local information of focus cloud due to its full connectivity and proposes an adaptive dynamic graph attention network.According to the input data,the network adaptively adjusts the graph structure dynamically to extract the local features of point cloud which are beneficial to registration,without manually setting the size of local area range.The algorithm is validated on the KITTI dataset and real scenarios including large indoor parking lots and open road environments.