Research On 3D Laser SLAM Based On Point Cloud Segmentation And Clustering

In recent years,intelligent warehouse handling and intelligent unmanned inspection have become the development trend of robotic industrial applications.SLAM(Simultaneous Localization and Mapping)is a prerequisite for robotic mobility in industrial applications.However,dynamic changes in objects in outdoor applications can easily lead to errors in the process of SLAM.Meanwhile,the 2D Li DAR commonly used in SLAM has too limited a description of the environment,which makes robots more likely to make errors in SLAM,thus limiting outdoor applications.To address these issues,this thesis focuses on segmentation and clustering of point clouds under 3D Li DAR SLAM,using accurate segmentation of ground point clouds and accurate clustering of non-ground point clouds to overcome the adverse effects of object dynamics and avoid errors in SLAM for mobile robots.The main work contents are as follows:(1)To address the problem of inaccurate ground segmentation,which leads to poor positioning accuracy and affects the real-time performance of mobile robots,an algorithm for ground point cloud segmentation based on concentric circle model is proposed.Firstly,concentric circles are constructed with sensor as the centre of circle,depending on the size of radius,and the concentric circles are divided into sectors according to different angular ranges to store the point cloud data.Then,PCA(Principal Component Analysis)is then used to fit a plane to the point cloud data within the sector,set thresholds to delineate the ground points,and perform adaptive updates to the thresholds.Finally,images are constructed based on the vertical and horizontal angles of the Li DAR,and the set of boundary points is obtained by image dilation,combined with Euclidean distances to construct convolution kernels to refine the boundary points as ground points or non-ground points.Ground segmentation experiments show that the algorithm performs well overall,maintaining a high level of precision,recall and accuracy.The algorithm maintains a consumption time of around 0.0582 seconds,which meets real-time requirements.(2)To address the problem of inaccurate clustering of non-ground point cloud data in inter-frame mismatching,a non-ground point cloud clustering algorithm based on harness relation is proposed on the basis of(1).Firstly,an initial clustering set is obtained by processing the intermediate scan line,and a kd(k-dimensional)tree is constructed with the points contained in this initial clustering set.Subsequently,diffusion to both sides is performed to cluster the points that meet the requirements in the kd tree,and similar classes are merged after processing adjacent scan lines.Then,corner and plane points are extracted from non-ground and downsampled ground points respectively by using different curvature thresholds set.Finally,corresponding corner and plane points of current frame are found in the same class of previous frames,and relative transformation between neighbouring frames is calculated by using the LM(Levenberg-Marquardt).Non-ground point clustering experiments show that the algorithm can achieve relatively high clustering accuracy and consume less time,staying around 0.02 seconds,which can meet real-time requirements.(3)Based on the ground point cloud segmentation algorithm and the non-ground point cloud clustering algorithm,an improved SLAM scheme for mobile robots is proposed.Then,an experimental platform is built to verify that the scheme was able to reduce positioning errors from an average of 3.2 metres to 1.3 metres and increase the clarity of the maps.