| The 3D laser scanning technology has the advantages of a comprehensive,efficient and accurate expression of the 3D information of objects.In recent years,with the continuous improvement of laser scanning equipment in performance and cost,3D laser point cloud processing algorithms have been applied to more and more scenarios,such as 3D reconstruction,SLAM(Simultaneous Localization and Mapping),autonomous driving,part detection,and reverse engineering.For example,in laser SLAM technology,the raw point cloud output by the laser radar has problems such as large data volume,uneven density distribution,noise,and incomplete information due to occlusion,which can cause difficulties for subsequent algorithm processing such as ground segmentation and registration.In addition,in laser SLAM mapping,data association is required to achieve pose estimation and incremental mapping,and point cloud registration is the core technology,which has a significant impact on the performance of SLAM mapping.Therefore,research on laser point cloud preprocessing and registration technology is of great significance.This article is based on the open source datasets LCAS and KITTI for indoor and outdoor autonomous driving,as well as a dataset collected at the Xianlin campus of Nanjing Normal University,including scenes from the campus underground parking lot and campus roads.The article delves into preprocessing and registration optimization methods for 3D laser point clouds.The research is divided into two parts: 3D laser point cloud preprocessing and 3D laser point cloud registration optimization.(1)To address the problem that the traditional point cloud voxel down-sampling requires manual adjustment of the raster size,this paper proposes a voxel down-sampling based on point cloud scale adaptation;To address the traditional statistical filtering is not able to remove the noise from the detail part near the plane,this paper proposes a statistical filtering algorithm for plane feature improvement.The experimental results show that this algorithm can greatly reduce the amount of data and retain most of the feature point clouds while filtering.(2)For the traditional ground segmentation algorithms mostly have the problems of low segmentation efficiency and insufficient segmentation,this paper studies the latest ground segmentation algorithm Patchwork++.The experimental results show that Patchwork++ has accurate and fast ground point segmentation efficiency,based on which the processing speed of point cloud alignment algorithm can be effectively improved.(3)The traditional point cloud alignment algorithm is difficult to perform fast and high accuracy alignment for point clouds with low overlap rate,so this paper improves the traditional RANSAC-ICP alignment framework and proposes an improved fast point cloud alignment algorithm.First,the key points are extracted from the source and target point clouds by combining the feature extraction algorithm of Intrinsic Shape Signatures(ISS),and then the key points are characterized by the Signature of Histogram of Orientation(SHOT)to obtain the high-dimensional In the coarse alignment stage of Random Sample Consensus(RANSAC),the SHOT feature descriptors of these ISS key points are used to initially obtain the sampled point pairs,add geometric constraints to eliminate the wrong point pairs,and obtain better initial poses;finally,the coarse aligned point cloud is iteratively aligned with the objective function symmetry.Finally,the coarse-aligned point cloud is finely aligned using the Iterative Closest Point(ICP)algorithm with symmetric objective function to further reduce the position difference between the two frames and complete the final alignment.The results of the alignment experiments show that the method generally outperforms newer algorithms in terms of average alignment speed and accuracy at low overlap rates.Finally,the algorithm of this paper is incorporated into the laser SLAM building process,and the actual building has achieved good results. |
PDF Full Text Request