Study On 3D Point Cloud Registration And Fusion

With the rapid development of point cloud acquisition technology,the study on the related technology of 3D point cloud has important research value,and recently has become a focus of scholar attention at home and abroad.The technology of point cloud registration and fusion is the basic research direction of 3D computer vision.It is widely used in 3D reconstruction,remote sensing analysis,navigation and mapping,and virtual and augmented reality.Real-world 3D point cloud have problems such as noise,outlier varying degrees of occlusion and uneven density,which bring great challenges to the fusion and reconstruction of 3D point cloud.Although this technology has achieved many progressive work,it is still a challenging issue.Focusing on their key issues.Point cloud registration and fusion involve key techniques such as local feature description,matching correspondence extraction,registration methods,and fusion.This topic has made the following progress:Aiming at the problems of occlusion,noise in the 3D point cloud in real world,a MultiScale Siamese Network(MSS-Net)is proposed to learn multi-scale local features.First,To obtain sufficient training samples,pairs of semi-synthesized 3D local patches are constructed to do data augmentation,which makes them more robust to rotation,noise,etc.Secondly,based on the varying local domain,a MSS-Net network is constructed to learn multi-scale local features.Finally,geometric constraints are introduced to remove mismatched correspondences.Experimental tests show that compared with representative methods,MSS-Net achieves the best performance.In addition,the multi-scale local features learned by MSS-Net are effectively applied to TLS point cloud registration.Aiming at the problem of high proportion of mismatched point pairs or outliers in point cloud registration,a method based on Supervoxel Guidance and Game Theory optimization(SGGT)is proposed to efficiently extract reliable matching correspondences or inliers,and applied for point cloud registration.First of all,in order to reduce the size of keypoint pairs,combining the three-dimensional spatial homomorphism of 3D supervoxels,the keypoint correspondences are grouped to obtain more powerful groups of keypoint correspondences.Secondly,in order to extract the promising combined groups,a ’fit-andremove’ strategy is proposed which incorporates the characteristics of three-dimensional space transformation,to achieve the coarse extraction of candidate matches.Finally,to extract higher-precision inliers,considering the relationships between the combined groups,a method of non-cooperative game optimization based on the combined groups is proposed to globally eliminate mismatched combined groups.Experimental results show that when the number of key point pairs increases(larger than 1,000),the SGGT algorithm is nearly 100 times faster than the state-of-the-art algorithm,while maintaining similar accuracy.Aiming at the problem that ICP and its derived methods,which base on singular value decomposition(SVD)for transform matrix,are susceptible to sampling points and noise,a registration method of data-driven self-supervised network(ST-Net)iterative optimization is proposed.First,a data-driven self-supervised transformation network is constructed to iteratively fit 6DoF parameters.Secondly,in each iteration,a Monte Carlo simulation method is constructed at random sample to improve the efficiency of point cloud registration.Finally,the ST-Net network is achieved robust registration of point clouds by iteratively minimizing the Hausdorff distance loss between pairwise point clouds.Comprehensive tests show that the ST-Net network iterative registration method is more robust,and less sensitive to the size of the point cloud.Aiming at the problem of layering and noise in 3D reconstruction of cross-source point clouds,a 3D volume fusion method based on improved graph cut model is proposed.Due to the differences in non-rigid deformation,occlusion,and density distribution of cross-source point clouds,there are problems such as layered redundancy and non-smooth gap boundaries in 3D reconstruction after registration.Therefore,first,to eliminate the layering problem of fusion,a graph cut model with boundary constraints is constructed to achieve the primary fusion of cross-source point clouds.Secondly,in order to smooth the fusion results,a gradual migration method is constructed based on the local average of the normal vectors.Finally,experimental results prove the effectiveness of the proposed algorithm in eliminating stratification.In summary,in terms of point cloud registration,this thesis proposes the solutions from different perspectives,such as multi-scale local features,match extraction method,and regression optimization of transformation parameter.On the fusion and reconstruction of cross-source point clouds,a method based on improved graph-cut model is proposed.The study of this PhD project has played an important role in promoting the research of 3D point cloud registration,and provided a new solution for multi-source 3D point cloud fusion and reconstruction.