Underwater SLAM Front-End Matching Method Based On Point Cloud Data

There are many secrets waiting for human beings to explore and understand in the ocean.In the environment perception problem,a common technology is SLAM,which integrates positioning and mapping,and can well assist carriers in exploring the environment.As a branch of SLAM,underwater SLAM has different difficulties from traditional SLAM.First of all,the general terrestrial and perceptual sensors cannot work well underwater and are easily restricted by the environment and noise.However,sensors used to perceive the environment in the water generally have conflicts between information richness and real-time.Faced with this dilemma,this paper proposes to use 3D real-time imaging sonar to provide data sources,and to initially solve the front-end problem of underwater SLAM by analyzing 3D point cloud data.The main content of the paper is as follows:Firstly,the data format of the point cloud and the information related to 3D real-time imaging sonar are introduced;for the concept of point cloud neighborhood,three common neighborhood search algorithms are listed and analyzed;for the problem of data redundancy,a mean filtering method is introduced;for the problem of noise in the point cloud,an adaptive point cloud denoising method is proposed and compared with other two classical denoising algorithms and the rationality of the designed denoising algorithm is confirmed by the experimental results;in addition,the calculation method of point cloud surface normal vector is also introduced.Secondly,two common point cloud feature point extraction algorithms are introduced:ISS and Harris3D,and a brief analysis of these two algorithms is presented;for deep learning techniques,three common ideas related to the use of deep learning to process point clouds are summarized,and the limitations of these ideas are pointed out;in order to make local point cloud data analysis more convenient,graph convolution neural networks are introduced,and two mainstream graph convolution methods are presented:spectral domain graph convolution and spatial domain graph convolution.Then,a novel network using unsupervised learning for feature point extraction is proposed based on the characteristics of traditional feature point extraction algorithms and the advantages of graph convolutional neural networks;the main idea of designing this network is presented;the specific structure and operation principles of the network are described in detail in two main parts(the feature point extraction part and the constructive loss function part);the loss function value curves of the proposed feature point extraction network are suitably analyzed,and the reasonableness of the algorithm in this paper is verified by comparison tests with the ISS algorithm and three types of point cloud data.Finally,the classical FPFH algorithm is introduced,its characteristics are also analyzed,and improvements are made to address the shortcomings in FPFH,weakening the dependence of FPFH on normal vectors and designing a more descriptive algorithm;to solve the problem of strong noise during inter-frame matching,RANSAC and ICP are combined,in which RANSAC overcomes the problem of much noise well and obtains a better initial value,while ICP achieves further matching of point cloud data,completing the task of SLAM front-end matching and initially confirming the rationality of the SLAM front-end matching process adopted in this paper.