Research On LiDAR Based Autonomous Localization For Mobile Robots

Technological advances are constantly changing the way people produce and live,and the fourth scientific and technological revolution,represented by artificial intelligence,is coming quietly.In the past,the robots,which embody many scientific and technological achievements,were mainly used in industrial production,but with the rapid development of robotics,mobile robots are gradually entering into people's daily lives.To this end,mobile robots need to be empowered to interact freely with people and the environment,and thus to increase their autonomy and intelligence,which depends on the various types of sensors they are equipped with.As one of the most advanced environment perception sensors at present,LiDAR has been developing rapidly in recent years and has undergone multiple rounds of technological updates.Because of its advantages of high measurement accuracy,high scanning frequency,large field of view,and robustness against varying environmental conditions,LiDAR can provide powerful environment perception capability for mobile robots and has been increasingly used in a variety of mobile robot platforms.For mobile robots,the prerequisite and basis for achieving true autonomy is the ability to localize themselves accurately in the operating environment.LiDAR based localization of mobile robots is currently an active research area,which is the core topic of this thesis.Several key technologies are studied in this thesis,including global scan matching,real-time LiDAR odometry,and LiDAR/IMU integrated localization.The main work and innovations are as follows:1.A global scan matching method based on planar feature is proposed to address the problem of low efficiency and accuracy of existing global scan matching methods.An incremental principal component analysis algorithm is adopted to improve the efficiency of planar feature segmentation of unordered point clouds;a point cloud plane fitting algorithm that takes into account sensor measurement noise is developed to improve the estimation accuracy of plane parameters;plane attributes and relationships between planes are comprehensively used for fast plane matching,which effectively avoids plane mismatching;coordinate normalization processing is introduced to improve the estimation accuracy of pose transformation,and a consistency metric based on the distance between the centroid and the corresponding plane is designed to select the optimal solution of pose transformation.The experimental results show that the coordinate normalization operation helps to improve the accuracy of transform estimation based on planar feature;the comparative experimental results demonstrate that the proposed method outperforms the existing methods in terms of accuracy,efficiency and robustness.2.A LiDAR odometry method based on key-points matching and continuous-time trajectory estimation is proposed considering that the existing LiDAR odometry methods cannot effectively handle the distortion problem induced by high-speed motion.A continuous-time trajectory estimation model on SE(3)is introduced,where Gauss process prior is incorporated to participate in the optimization solution.Distortion correction and trajectory esitamation are performed simultaneously in a unified optimization framework,and Gauss process interpolation is used to query the pose and velocity at any given time on the trajectory;after efficient ground points detection,the extraction of key-points is realized using range-corrected intensity and curvature together.Sliding window optimization is utilized to take full advantage of the overlapping constraints between consecutive multiple frames;a robust kernel function is introduced to improve the robustness against outliers,and an adaptive strategy is adopted to determine the node interval.The experimental results show that the extraction of key-points and the use of overlapping information between multiple frames are beneficial to improve the accuracy of motion estimation;the speed of motion has a relatively large impact on the accuracy of LiDAR odometry,but continuous-time trajectory estimation can be used to better deal with this problem;the proposed method is superior to the comparative methods in terms of localization accuracy and robustness,and the average relative accuracy can reach less than 1% when applied to the KITTI dataset.3.In view of the fact that there are few fast and effective plane segmentation methods for multi-beam LiDAR point clouds,an efficient plane segmentation method considering the multibeam LiDAR measurement characteristics and scanline structure is proposed.Spherical distance image projection is used to transform the original scan into an ordered point cloud.After recovering the scanline structure of the scanned point cloud,each scanline is processed individually.Specifically,the edge points in each scanline are determined through depth and surface direction continuity assumption,and then are used to divide each scanline into subsegments;a plane detection technique similar to Hough transform is performed on each merged sub-segment.After plane filtering,the final planar features are determined by splitting-merging,and the results are further refined using point-based region growing.The experimental results indicate that the proposed method can achieve fast and effective plane segmentation of multi-beam LiDAR scanned point clouds,and the processing time of a single frame is about 0.04 s,which meets the requirements of real-time processing.4.To further improve the localization accuracy and efficiency of LiDAR odometry in structured or semi-structured scenarios,a LiDAR odometry method based on hybrid scan matching and plane landmarks is proposed.In the hybrid scan matching method,two planes from adjacent frames are accepted to be identical if they are approximately parallel and close to each other.A constraint ellipsoid parameterized by a constraint matrix is used for constraint analysis.When the constraints provided by plane correspondences are sufficient,the transform estimation is directly performed based on the planar features.Otherwise,minimum number of key-points are selected and are combined with planar features to estimate the pose transformation through an iterative optimization manner similar to the NICP algorithm.The odometry factor obtained by scan matching and planar feature observation factor are both utilized to connect nodes,which represent poses and plane parameters,to construct a factor graph model for optimizing the states to be estimated.The experimental results validate that the proposed method can effectively reduce the cumulative localization error,and can achieve higher accuracy than LOAM and GICP while maintaining real-time processing.5.LiDAR-only based localization has relatively low update rate,and is not robust to rapid motion and large rotation,thus leading to large accumulated error.Considering the insufficient fusion nature of existing integrated localization methods,a LiDAR/IMU tightly coupled localization method is proposed by using a factor graph optimization framework,where four factors including inter-frame motion constraints,plane landmarks observation constraints,IMU pre-integration constraints and loop closure constraints are cooperated to perform online incremental optimization.Extrinsic parameters between LiDAR and IMU are obtained by means of hand-eye calibration.Edge and plane points are extracted using difference of neighboring point normals;point correspondences and plane correspondences are combined to estimate frame-toframe motion;the pre-integration technique is adopted to solve the rate inconsistency problem between LiDAR and IMU,and the pre-integrated results are adequately used for motion compensation and scan matching.The local and global consistency is further enhanced by mapping and loop detecting modules.The experimental results demonstrate that the proposed LiDAR/IMU integrated localization method can achieve higher accuracy than comparative methods across different kinds of scenarios.Taking trajectory closing error as metric,the proposed method can achieve decimeter or centimeter accuracy.In addition,the point cloud map built through real-time localization shows better local and global consistency.