Simultaneous Localization And Mapping Algorithms For Indoor Robots Based On Three-Dimensional Laser Point Cloud

Simultaneous Localization and Mapping technology refers to the incremental map creation of the environment by moving of robot from an unknown location in an unknown environment,and completing the robot location and navigation independently via the map created.SLAM technology can be divided into visual SLAM and laser SLAM according to the types of sensors used.Among them,laser SLAM has the characteristics of high precision,fast speed and quick response to dynamic environment changes.Therefore,the reliability and safety of laser SLAM technology are superior to visual SLAM.Laser SLAM technology is widely used in self-driving cars,robotics,indoor navigation and positioning,so this topic not only has a high economic and social value,but also has a high signific ance in academic research.This paper mainly studies SLAM problem of indoor autonomous mobile robot based on laser,including four parts: data preprocessing of laser point cloud,point cloud registration,loop detection and complete experimental verificati on of laser SLAM.Specifically:Firstly,in order to eliminate most outliers effectively,three traditional point cloud filtering methods,voxel filtering,statistical filtering and radius filtering,are compared and analyzed,and a rough processing scheme with statistical filtering as the main and other two filtering as the supplement is determined.At the same time,in order to smoothen the point cloud effectively and keep the sharp edge feature of the point cloud,a point cloud denoising method based on L0 minimization is proposed.We verify the effectiveness of our method by simulation experiments.Secondly,aiming at the problem that the current point cloud registration algorithm does not consider the local surface and geometric information of the point cloud,a coarse-to-fine 3D point cloud registration algorithm based on affine invariant ratio is proposed.The distance weights of SHOT descriptors are modified to reduce the influence of noise points in the neighborhood of feature points on descriptors.Aiming at the omission of local geometric shape information of point clouds and the lack of inaccurate point pairs' elimination in current registration algorithms,a feature point constraint based on affine invariant ratio is proposed,which introduces local surface and geometric information of point clouds and effectively eliminates mismatches.Aiming at the large cumulative error of continuous multi-frame point cloud registration,a neighbor iteration registration strategy is proposed.Thirdly,aiming at the problem of large amount of computation in histogram looping detection method,a looping detection algorithm based on point cloud appearance descriptor is proposed.By adjusting threshold,the appearance descriptor is divided into subclasses,the descriptiveness of the appearance descriptor is increased,and the computation amount of the algorithm is adjusted.The point cloud is rotated to ensure the rotation invariance of its appearance descriptor.Aiming at the problem that similar scenes are prone to false-positive detection problems,the constraints of distance information are introduced to reduce false positive loop judgment.Finally,the improved algorithm is integrated into the existing SLAM framework Cartographer,and its effectiveness is verified by experiments.The experimental results and comparative analysis of point cloud preprocessing algorithm,point cloud registration algorithm and loop detection algorithm are carried out respectively.The validity of the improved SLAM algorithm is verified.