Research On Unsupervised 3D Point Cloud Registration Algorithm Based On Deep Learning

In two-dimensional computer vision,deep learning has made breakthroughs in many domains and has outperformed traditional computer vision methods in many areas.Point cloud registration is a 3D extension of the image matching problem in 2D computer vision.Currently,there are two main types of algorithms for point cloud registration,namely,transform estimation algorithms based on the global distribution and feature matching algorithms based on local feature extraction,where the former directly estimates the transform parameters between point cloud pairs based on the global distribution and the latter is based on several steps:key point detection,feature extraction,feature matching and transform estimation.In this paper,deep learning methods are introduced to the field of 3D vision,and an unsupervised point cloud registration algorithm based on local feature extraction is proposed.The proposed algorithm includes a data preprocessing method based on classical algorithm,a point cloud partition strategy based on adaptive extended receiving domain,a key points extraction network combined with attention mechanism,a local feature extraction network based on hierarchical information fusion,and a point cloud transformation estimation algorithm based on the RANSAC algorithm.Compared with other current algorithms,this algorithm finally achieved excellent registration accuracy with a faster registration speed.More importantly,this algorithm is an unsupervised algorithm.The research work of this topic is mainly as follows.(1)In terms of point cloud key points extraction,an unsupervised key points extraction algorithm based on adaptive extended receiving domain point cloud partition strategy is proposed.Firstly,the distribution of point features in the point cloud is introduced,the input data enhancement method based on normal calculation and the interference points removal method based on ground segmentation are introduced.Then a point cloud partition strategy based on adaptive extended receiving domain and a key point extraction network based on attention mechanism are proposed and an unsupervised training network with an efficient loss function is given.Finally,experiments are conducted on the widely used datasets in the field of point clouds,and the experimental results are analyzed to demonstrate the effectiveness of the unsupervised local point cloud feature extraction algorithm.(2)In terms of local point cloud feature extraction,an unsupervised local point cloud feature extraction algorithm based on hierarchical information fusion is proposed.Firstly,the characteristics of effective local point cloud feature are introduced,and a feature extraction network based on hierarchical information fusion is proposed.Then an unsupervised training network with a loss function fusing key point spatial information and uncertain saliency weights is introduced.Finally,experiments are conducted on common point cloud registration datasets and the results proved the effectiveness of the unsupervised local point cloud feature extraction algorithm.(3)In terms of feature matching and transformation estimation,a RANSAC-based algorithm for point cloud transformation estimation is proposed.The specific design of each stage of the algorithm used to improve the computational efficiency and robustness is described in detail.Finally,the proposed algorithm is validated with experiments and the experimental results are analyzed to demonstrate the effectiveness of this point cloud transformation estimation algorithm.