| 3D point cloud registration plays an important role in 3D reconstruction,automatic driving,surveying and mapping and other fields.Traditional ICP and other registration algorithms are easy to fall into local optimization and are sensitive to noise.With the increasingly widespread use of point cloud data and the rapid development of deep learning technology,some studies gradually use deep learning algorithms for point cloud registration,but there are still problems that the registration scheme degrades when the point clouds overlap large cloud registration error of large scenic spots.This paper uses deep learning algorithm to extract point cloud feature descriptors for point cloud registration,and constructs a laser odometer scheme based on deep learning.The existing deep learning point cloud feature description sub network is optimized to build a deep learning network suitable for point cloud registration.By analyzing the data and scene,the registration features are constrained to improve the point cloud registration performance when the point cloud partially overlaps and the performance in the laser odometer.The main work of this paper is as follows.Firstly,this paper summarizes the research background and significance of this topic,analyzes the current situation of point cloud feature descriptor extraction and point cloud registration at home and abroad,analyzes the existing problems of point cloud registration,introduces some point cloud data processing algorithms and point cloud feature descriptor extraction algorithms,and formulates the main research objectives and directions of this paper.Secondly,aiming at the problems of the existing feature descriptor extraction network based on deep learning,such as the lack of local feature information and global feature information in the process of network propagation,a multi-level feature fusion structure with self attention mechanism is proposed,so that the obtained point cloud feature descriptor has local feature information,global feature information and spatial context information at the same time.The effectiveness of this method for registration is proved by experimental analysis on public data.Thirdly,aiming at the problem of registration degradation in the existing partial registration methods when the point clouds are partially overlapped,especially in the outdoor scene with multiple noise scenes,a feature descriptor extraction network based on multi feature cross mapping fusion is proposed.Through the point cloud transformer structure,a cross mapping structure based on graph convolution is constructed,and the point correspondence is obtained according to the principle of feature similarity to increase the registration efficiency.Through the experimental analysis in the data set and the comparison with the registration without adding the modified module,the effectiveness of this method is proved.Finally,a laser odometer scheme based on deep learning is constructed for the outdoor scene.Through the analysis of point cloud data,a point cloud ground filtering method is designed to filter the ground point cloud while retaining the structure of the remaining point cloud.In addition,by analyzing the foreign environment and data acquisition,a loss scheme suitable for the laser odometer scheme is designed to restrict the negative correspondence of point cloud features and enhance the positive correspondence of point cloud features.The effectiveness of the scheme is proved by comparing the scheme with the classical odometer. |
