Research On 3D Point Cloud Registration Optimization Method Based On Laser SLAM

In recent years,the scanning application scenarios of multi-line LiDAR sensors have increased with the popularization of depth sensors and the development of 3D scanning technology,and Simultaneous Localization and Mapping(SLAM)technology based on multi-line LiDAR sensors has become the research hotspot.In 3D laser SLAM technology,data needs to be correlated to achieve pose estimation and incremental mapping,and 3D point cloud registration is the core technology in data association.There are some problems in the 3D laser point cloud registration,for example,the 3D laser point cloud has large amount of data,uneven density distribution,complex noise points with complex scenes and missed data due to occlusion.In order to improve the efficiency of 3D laser point clo ud registration,this paper focuses on the 3D point cloud registration optimization method of laser SLAM.The main research object is 3D laser point cloud.The main research content is divided into three parts: 3D laser point cloud pretreatment,3D laser point cloud feature extraction and 3D laser point cloud registration optimization.Aiming at the ground points segmentation part of 3D laser point cloud preprocessing,this paper proposes a new method based on two-dimensional ranging image to segment the ground points.The experimental results show that the method has the more accurate and fast ground points segmentation efficiency and can be transplanted to real-time 3D laser SLAM algorithm performance.In the aspect of 3D laser point cloud feature extraction,this paper studies the ISS(Intrinsic Shape Signatures)feature extraction algorithm and Harris feature extraction algorithm,and proposes a Voxel-SIFT feature extraction algorithm.In the paper,three algorithms are compared,the experimental results of six different laser scanning scenarios are analyzed,and the effectiveness of Voxel-SIFT feature extraction algorithm is verified.In the aspect of 3D laser point cloud registration,this paper improves the iterative closest point registration algorithm combined with Voxel-SIFT feature extraction algorithm,and a coarse-fine fast point cloud registration method using Voxel-SIFT feature is proposed.The method first selects the feature points to coarse match,and then uses the random sample consensus method to eliminate the false matching points and optimizes the initial transformation parameters.Finally,the K-dimensional tree neighbor search method based on the optimized initial transformation parameters,combined with iterative closest point registration algorithm,is used to complete the fast and accurate registration of the original point cloud data.The test results of the six typical scene data from open source datasets and being collected in practice show that the average registration time of this method is 78% shorter than the traditional ICP registration algorithm.