Research On Key Technologies And Applications Of LiDAR SLAM Syste

Since the Fourth Industrial Revolution,robot intelligent platforms have become more widely used worldwide.It can be used in fields such as unmanned distribution,smart agriculture,and autonomous driving.As an important part of the unmanned platform,SLAM(Simultaneous Localization And Mapping,real-time positioning and mapping)system can ensure that the task of autonomous positioning in an unknown environment is completed,and also can incrementally build environmental maps.Therefore,the research on SLAM is of important practical significance.Using the Lidar,this thesis conducts in-depth research on the key technologies involved in the SLAM system.The specific work content is as follows:Firstly,the preprocessing of laser point cloud data is studied.For the elimination of motion distortion errors,this thesis uses a uniform motion model to compensate for the rapid distortion.For eliminating the point cloud noise and isolated points,a Radial Filter Based on Downsampling(RFBD)based on voxel downsampling is designed.The KITTI data set simulation then shows that this method can reduce the number of point clouds to about 1/3 of the initial point cloud.Moreover,while ensuring the filtering effect,the processing speed of this method is increased by 4.7 times compared with the traditional statistical filtering method.Moreover,the design of front-end odometer based on point cloud frame matching algorithm is studied,and a GSC-NDT-GICP fusion matching algorithm(NDT-GICP based on Ground Segmentation and Clustering)is proposed on the basis of NDT and ICP.Based on the ground segmentation and clustering of point cloud rasterization,this registration method can prevent ground points and dynamic objects from interfering with the registration.At the same time,the "coarse to fine" method from NDT-GICP registration is adopted,to improve the matching accuracy and avoid the complex iterative operation,thereby improving the operation efficiency of the registration.Comparative experiments in various environments of the KITTI data set has shown that the matching algorithm proposed in this thesis can reduce the translation error rate of positioning to 1.11%,and the average positioning accuracy is even 35.05%higher than that of the LOAM odometer in rural scenes.It is also verified that the odometer system still has good robustness at different speeds.Furthermore,based on the odometer,the loop detection and back-end global optimization problems are studied,a two-node structure loop detection algorithm thereby being proposed.The two nodes refer to the judgment node based on the position relationship and the verification node based on the similarity of the point cloud space descriptors.Simulation comparison results show that the detection algorithm in this thesis can increase the loop recall rate to about 65%under the premise of 100%accuracy.For the global optimization,based on the idea of PGO(Pose Graph Optimization),this thesis uses nonlinear least squares to solve the optimized pose,minimizing the global error.Compared with the simulation results of the odometer,the average accuracy of the system positioning is increased by 30.63%after adopting loop optimalization,effectively eliminating the accumulated drift error of the odometer.Finally,this thesis adopts laboratory unmanned vehicle as the research platform,to test and verify the system in multiple local scenes on campus.The results show that the SLAM system constructed in this thesis can better achieve the goal of real-time accurate positioning.At the same time,the constructed three-dimensional point cloud map highly restores the actual scene.