Research On The Key Technology Of 3D Point Cloud Registration Based On Depth Images

With the continuous development of depth sensor technology,the application of 3D point cloud has shifted from professional to daily consumer field.In order to obtain the complete depth image data of different scenes,multiple depth cameras are usually required to sample at different viewpoints.Obviously,registration becomes the most important part of point cloud data processing process.Due to the characteristics of detailed presentation,large data amount and high deformation probability,the smallscene point cloud data requires the rigid/non-rigid registration computation with high efficiency,high accuracy and practicality.The existing point cloud registration models have several problems:First,the registration methods based on point correspondence do not take into account the global characteristics of the local feature matching process or have insufficient point feature expression capability,which results in lower registration precision;Then,the practical applications of common non-rigid alignment models are limited due to low computational efficiency,weak generalization ability or complex modeling.To address the above issues,this dissertation focuses on the research of key technologies of small-scene point cloud data as follows:1.The research of rigid registration based on bipartite graphThrough extracting key points in the coarse process and computing them instead of original data in the fine phase,the registration efficiency can be improved.In order to reduce mismatching rate,we introduce a bipartite graph structure and treat the point correspondence as a graph assignment problem.Then,a registration strategy is designed to combine local features with global assignment.Moreover,we present a pluggable Top-Tail strategy to extend this registration model to anisotropic scaling scenarios with a small amount of computation.Compared with other similar methods,the experiment proves that the proposed model has highest registration accuracy in data scenarios with missing data,low overlap rate and inconspicuous features.2.The research of rigid registration based on graph and deep learning frameworkTo address the point correspondence mismatch caused by insufficient feature representation,this dissertation combines the graph structure and neural network to implement two registration models that fully embed global and local features.In order to enhance the description of point cloud features,the proposed dual graph matching registration model incorporates irregular shape factors into spatial,global and topological features.The hierarchical probabilistic distribution registration model utilizes the local to global spatial distribution for feature expression.Moreover,the noise can be filtered by an intra-graph information interaction mechanism and missing points mismatching can be avoided by cross-graph rotational consistency calculation.The experiments verify that the dual graph matching registration model has SOTA performance dealing with complete,noisy and incomplete data(the lowest registration error close to 0%);the hierarchical probability distribution registration model is competitive in missing and low overlap rate data scenes.3.The research of non-rigid registration based on unsupervised learningFor depth image registration with large deformation,inspired by the inverse computation of Lucas-Kanade algorithm and recurrent neural network state update process,this dissertation transforms the non-rigid registration problem into a series of rigid transformations and proposes an unsupervised learning non-rigid registration model.To eliminate the effect of invalid rigid transformations,the attention mechanism is utilized to determine the degree of retention in different rigid transformations.In order to improve the registration accuracy,based on the 2D manifold representation and reconstructed point cloud,we design a loss function for the deformed shape similarity measure;Then,with the spatial probability distribution model,we implement a loss function to measure the neighborhood structure consistency,which remains the point local structure unchanged.Compared with other similar nonrigid registration models,this method has better performance both in testing dataset and real sampled data.