Research On SLAM Algorithm Based On LiDAR

The rapid development of driverless technology is transforming various aspects of people’s lives,such as unmanned delivery,cargo operations at ports,and Robotaxi.In the localization and planning phase of driverless driving,high-precision road network maps play a crucial role,which are generated based on environmental maps constructed using Simultaneous Localization and Mapping(SLAM)algorithms.The current mainstream laser SLAM algorithm employs a mechanical 3D LiDAR sensor,which is known for its high accuracy but comes with a significant cost,making it challenging to use for mass production of unmanned vehicles.With recent advancements in hardware technology,solid-state LiDAR with point cloud density up to image level has become available in the market.Compared to traditional mechanical LIDAR,solid-state LIDAR point clouds offer more detailed information about the environment and come with a lower price.However,using solid-state LIDAR also presents certain challenges such as non-repetitive scanning,lack of ring information in the point cloud,and a narrower maximum view angle of 120°,leading to blind spots in the view.To overcome these challenges and utilize solid-state LIDAR for unmanned vehicle localization and map building,this paper investigates a set of SLAM algorithms specifically designed for solid-state LIDAR.The key contributions of this study are outlined as follows.(1)This paper presents a feature extraction algorithm based on surface topography.The proposed method utilizes a normal distribution probability density function to calculate the sampling density and downsamples the point cloud.Then,the point cloud morphology features are extracted using principal component analysis due to the absence of ring information in solid-state LIDAR.(2)This paper presents a loop closure detection algorithm based on a doublechecking mechanism.The proposed algorithm utilizes global descriptors that match the characteristics of solid-state LIDAR view generated through region segmentation of Scan Context.Loop closure matching pairs are obtained by calculating the similarity of the descriptors.The matching pairs are then verified twice using NDT(Normal Distributions Transform)alignment.(3)This paper presents an improved factor graph-based global pose optimization algorithm.The proposed algorithm constructs a graph model that includes odometry factor,preintegration factor,and loop closure detection factor.Additionally,a keyframe selection strategy is designed to dynamically adjust the joining frequency of the odometry factor in the graph model based on the change in pose between adjacent frames.Furthermore,this paper presents a campus road dataset constructed using sensors and data acquisition platforms,including solid-state LiDAR and mechanical LiDAR provided by the laboratory.The dataset is used to validate the proposed SLAM algorithm based on solid-state LiDAR and conduct comparative experiments between the proposed method and mainstream mechanical LiDAR SLAM algorithms.