Research On Multi-Dimensional Robot State Sensing With LiDARs

LiDARs are widely used in sweeping robots,unmanned vehicles,and various service robots due to their accurate,efficient,and stable spatial information acquisition capabilities.Point cloud is a typical representation of LiDAR data.Recently,robot perception and localization algorithms based on LiDAR point clouds have received extensive attention from academia and industry.Mature laser perception algorithms can quickly promote the industrialization of service robot products.For example,laser based simultaneous localization and mapping algorithms help sweeping robots enter thousands of families.Although the current research in the field of laser perception has made significant progress,the increasingly complex tasks and dynamic and changeable working environment pose severe challenges to the robustness and real-time performance of laser perception algorithms.On the one hand,service robots are primarily deployed in large-scale and high-dynamic places,such as restaurants,hotels,shopping malls,hospitals,etc.Factors such as object occlusion,light changes,feature loss,scene degradation,and frequent start-and-stop will cause the performance of the robot perception and localization algorithm to decline or even fail.On the other hand,due to the limitation of the computing resources of the service robots,the improvement of the robot’s work intensity has higher requirements for the real-time performance of the LiDAR perception algorithm.Moreover,traditional LiDAR perception algorithms remain in analyzing primary geometric features and lack advanced reasoning on those features.LiDAR perception algorithms urgently need to move towards the era of spatial intelligence.Given the above problems and challenges,this paper studies how to effectively utilize the advanced geometric information in LiDAR point clouds to improve the efficiency,accuracy,and robustness of robot perception and localization algorithms.The LiDAR sensors involved in this paper include one-dimensional laser rangefinders,2D laser scanners,and spinning multi-beam LiDARs.The main research contents and contributions are as follows:1.For one-dimensional laser rangefinders,the data format is a point on the positive half-axis of the number axis,which can be called a 1D laser point or 1D point cloud.Inspired by the Cavendish experiment,this paper establishes the quantitative relationship between the robot joint angle and the laser traveling distance through planar geometry modeling,and proposes a flexible angle measurement method: "dual laser goniometric method".According to this principle,an angular sensor named "dual laser goniometer" is designed.Compared with traditional angle sensors,such as optical encoders,this device is not limited by the installation position,and can meet the needs of various types of angle measurement such as shaft joints,and soft robots.The working principle of the sensor is discussed in detail,and the model generalization in 3D space is carried out based on the Cartesian coordinates,which proves the unity of the angle measurement equation in 2D space and 3D space.In order to improve the angle measurement accuracy,a sensor intrinsic calibration method is proposed by the nonlinear least squares,which effectively reduces the angle measurement error.Exhaustive physical experiments show that the proposed dual laser goniometer can achieve high-speed and high-precision joint angle measurement.2.For 2D laser scanners,the data format is a point set in 2D Euclidean space,which is generally called 2D laser point cloud or single-line point cloud.In this work,the idea of plane geometry modeling is continued,and the single plane model is extended to a multi-plane model for 2D point clouds.An analytical planar laser odometry method named "Circle Fit Matching" is proposed.The algorithm takes plane features as input,which are ubiquitous in the structured indoor environment.In view of the working space of service robots,plane feature are mostly walls and furniture surfaces,which can be stable for a long time.Those stable features effectively overcomes the challenges of dynamic crowds and feature loss.Furthermore,a degeneration handling method based on the continuous motion assumption is proposed to weaken the influence of degraded scenes such as long corridors.The experimental results show that this analytical planar laser odometry algorithm has strong robustness and high efficiency.3.For spinning multi-beam LiDARs,the most common data representation is 3D point cloud,also known as a multi-line point cloud.In this work,an efficient plane segmentation algorithm based on line segment clustering is proposed to extract the planar structure information in 3D point cloud.This algorithm utilizes the spherical imaging model of multi-beam LiDARs,and uses radial projection and vertical projection to decompose the 3D point cloud into two sets of organized2 D point clouds.After single-line extraction on single-line point cloud,a planar region growing algorithm based on greedy breadth-first search is proposed.The experimental results show that the method performs well in both sparse and dense point clouds,and the running speed exceeds the current mainstream multi-plane segmentation algorithm.4.On the basis of 3D plane feature extraction,the single-line point cloud registration method proposed in 2D space is extended in 3D space,and a "Sphere Fit Matching" algorithm for multi-line point cloud is proposed,which derives an analytical solution for the 6-DOF robot motion estimation.In order to quickly establish effective matching of 3D plane features,a two-stage plane matching method in 3D space is proposed,which can accurately and efficiently establish correspondences between two groups of 3D plane features.Furthermore,in order to reduce the interference of the plane fitting noise,the robustness of the analytical motion estimation formulas have been improved by the weighted least squares and L1 norm optimization methods.The experimental results show that this algorithm can obtain good results of robot 6-DOF pose estimation in structured environments,and the running speed of this algorithm is fast,which can meet the real-time state estimation requirements of service robots in 3D space.Starting from the plane geometry analysis of low-dimensional LiDAR data,this paper gradually extends the plane geometry modeling method to 2D point cloud and 3D point cloud,and solves the robot single-joint rotation,3-DOF planar motion,and 6-DOF motion,respectively.A unified geometric robot state estimation theory is presented.In view of the above research contents and achievements,the proposed algorithms are verified and evaluated through detailed experiments.The experimental results fully demonstrate the effectiveness,efficiency and robustness of the proposed robot state estimation method.The innovative exploration of this paper will promote the further improvement of service robot technology,and also provide certain inspiration for the following robotics research.