| The objective of 3D point cloud registration is to achieve spatial alignment of point clouds by utilizing geometric or semantic features.It finds extensive applications in fields such as remote sensing,3D reconstruction,simultaneous localization and mapping(SLAM),etc.The registration process involves two distinct tasks based on different input data: "model-to-scene" registration and "scene-to-scene" registration.However,these tasks encounter several significant challenges: 1)in model-to-scene registration,the substantial differences in resolution and noise levels between the model and scene adversely affect descriptor stability,posing difficulties in establishing robust geometric correspondences;2)high-discriminative descriptors often exhibit high feature dimensions and lack compactness,thereby failing to meet the real-time registration requirements of resource-limited platforms;3)efficient construction of stable point correspondences between scenes becomes challenging due to the vast amount of 3D scene sequence data and uneven point cloud distribution;4)the presence of moving objects in open scenes causes significant interference in estimating global geometric consistency within the scene.In response to the aforementioned challenges,this dissertation conducts the following research:To address the high-precision registration requirements of model-to-scene scenarios,a high-precision point cloud registration method based on the Triple Multiple Attribute Correlated Statistics(Tri MACS)feature is developed.This method extracts resolutionindependent local surfaces,establishes a reference frame through locally stable axis voting,and describes features related to stable axis angle,tangential plane distance,and axial distance co-occurrence distributions.By iteratively estimating rotation and translation parameters,this method achieves high-precision point cloud registration,even in the presence of significant resolution differences and strong noise interference.To fulfill real-time 3D registration requirements on resource-constrained platforms,a point cloud registration method based on the maximum clique and the Rotational Projected Binary Structure(RPBS)feature is proposed.This method comprehensively utilizes multiple-view silhouette representation,occupancy grid binary feature extraction,weighted Hamming distance pre-screening,and geometric consistency maximum clique precision screening.It significantly enhances registration efficiency while ensuring registration accuracy.To accomplish robust and efficient registration of large-scale and non-uniform scene sequences,a point cloud registration method based on the multiple level height projection map is introduced.This method integrates global reference surface extraction,projection of iso-surfaces,and weighted inter-layer matching aggregation strategies.Consequently,it achieves a dual improvement in both registration accuracy and efficiency.To address stable registration requirements in scenes with moving objects,a targetlevel point cloud registration method considering moving object interference is presented.This method incorporates reference target extraction based on instance segmentation,target triplet relationship description,and parameter estimation based on the maximum "target overlap ratio".These advancements lead to significant improvements in registration accuracy and effective identification of moving targets.In summary,this doctoral dissertation provides an in-depth study on key aspects of high-precision registration,algorithm efficiency improvement,large sparse scene reconstruction,and moving object elimination within the context of 3D point cloud registration.The proposed methods are rigorously validated using publicly available datasets and have been successfully applied to tasks such as space station attitude estimation and complex environment perception,resulting in expected outcomes that support major national needs. |
