文章：LiDAR-based SLAM for robotic mapping: state of the art and new frontiers
Author: Xiangdi Yue and Miaolei He
Editor: Point Cloud PCL
Welcome to join Knowledge Planet, get PDF papers, and forward them to your circle of friends. The article is for academic sharing only. If there is any infringement, please contact us and delete the article. Please do not reprint without the consent of the blogger.
The public account is dedicated to sharing content related to point cloud processing, SLAM, three-dimensional vision, high-precision maps and other fields. Everyone is welcome to join. If you are interested, please contact [email protected]. Please do not reprint without the author's permission. All students are welcome to actively share and communicate.
Summary
Over the past few decades, the field of robotic cartography has experienced extensive research and development in simultaneous localization and mapping (SLAM) technology based on LiDAR (Light Detection And Ranging, LiDAR). This article aims to provide an important reference for researchers and engineers in the field of robotic cartography. This article focuses on investigating the current research status of LiDAR-based SLAM technology in robotic cartography from the perspective of various LiDAR types and configurations. A comprehensive literature review is conducted on SLAM systems based on three different LiDAR forms and configurations. We conclude that multi-robot collaborative map construction and multi-source fusion SLAM systems based on 3D LiDAR, combined with deep learning, will be a new trend in the future. To the best of the authors' knowledge, this is the first in-depth investigation of robotic cartography from the perspective of various LiDAR types and configurations, and it can serve as a theoretical and practical guide for academic and industrial robotic cartography development.
introduce
The development history of LiDAR-based SLAM is shown in Figure 1. The development history of LiDAR-based SLAM. In the classic period, filtering methods were the main way to solve SLAM problems. In order to solve SLAM in a Bayesian network, the filtering algorithm must collect the information at each moment in real time and divide it into the probability distribution of the Bayesian network. This filtering method represents an online SLAM system that obviously incurs significant computational overhead and can only produce small-scale maps. For large-scale mapping, an optimization strategy has been proposed to solve SLAM in factor graphs. The optimization method is the inverse of the filtering method, which only accumulates the acquired information and calculates the robot's trajectory and waypoints offline using the global information accumulated in all previous instances. In other words, the optimization method is a complete SLAM system. With the substantial improvement in computer performance and mathematical capabilities, optimization-based methods have become the main focus of contemporary SLAM research.
Figure 1. Development history of LiDAR-based SLAM
Filtering methods are the main way to solve SLAM problems in the classic era. In order to solve the SLAM problem in a Bayesian network, the filtering algorithm must collect information at every moment in real time and divide it into the probability distribution of the Bayesian network. This filtering method represents an online SLAM system that obviously incurs significant computational overhead and can only generate maps within a small area. For large-scale mapping, an optimization strategy to solve SLAM in factor graphs has been proposed. The optimization method is the inverse of the filtering method, accumulating only acquired information and calculating the robot's trajectory and waypoints offline using the global information accumulated in all previous instances. In other words, the optimization method is a complete SLAM system. With the huge improvement in computer performance and mathematical capabilities, optimization-based methods have become the main focus of contemporary SLAM research. As shown in Figure 2, overview of LiDAR-based SLAM system.
Figure 2. Overview of LiDAR-based SLAM system
LiDAR odometer
The purpose of LiDAR odometry is to generate a local map by creating an estimate of motion between two adjacent point cloud frames. LiDAR odometry is divided into three types according to point cloud registration methods: point-based registration, point-based registration Distributed registration and feature-based registration.
Point-based point cloud registration finds the correspondence between the target and the reference point cloud in the most direct way. A simple method to identify the corresponding point in the reference point cloud is to find the point with the shortest Euclidean distance, that is, iterative The closest point. Besl proposes a general, representation-independent method for accurate and efficient registration of three-dimensional shapes based on the iterative closest point (ICP) algorithm, which determines the best point-to-point correspondence with the minimum square distance. Better Euclidean transformation.
Unlike point cloud registration, distribution-based registration methods transform the point cloud space into voxels with a continuous probability density function. Matching the continuous probability density function of the target point cloud to the reference point cloud optimizes the pose connection. In 2003, Biber first demonstrated an alternative representation of range scans called the Normal Distribution Transform (NDT). Similar to the occupancy raster, Biber divides the 2D plane into cells, and for each cell, a normal distribution is assigned, which locally models the probability of the measurement points. The concepts of ICP and NDT are based on direct registration of point clouds. This direct registration method is time-consuming and difficult to provide in real time. Therefore, Zhang proposed LOAM to perform point cloud registration of two adjacent frames, which improves the efficiency of the system and enhances the registration accuracy by extracting geometric features.
Loop closure detection
Global data association generates a globally consistent map to correct accumulated errors by identifying whether the robot has reached the position it reached at the historical instant. Loop closure detection usually has two steps: (1) Use position recognition to find similar observations to the current observation. points in the database. (2) Pose graph optimization corrects the estimated loop pose. Methods for detecting loop closure using LiDAR can be further divided into two methods: local descriptor-based and global descriptor-based.
Graph optimization
The cumulative error of LiDAR makes long-term mapping inaccurate, but LiDAR odometry can quickly generate trajectories and maps. So in order to use LiDAR odometry to determine ideal routes and maps over long periods of time, a large-scale optimization problem must be created. Graph optimization is a method that achieves overall optimization by combining the pose and inter-frame motion constraints of each radar frame. It helps eliminate local accumulated errors in large-scale mapping and optimize trajectories.
LiDAR-based SLAM system
We will conduct a comprehensive literature review of LiDAR-based SLAM systems based on three different LiDAR forms and configurations, including (1) 2D LiDAR-based SLAM systems; (2) 3D LiDAR-based SLAM systems; and (3) rotation-based SLAM systems. shiLiDAR’s SLAM system.
SLAM system based on 2D LiDAR
Single-line LiDAR consists of a single-line laser module and a rotating mechanism. The scanning point of single-line LiDAR is usually within 360 degrees of the same plane, that is, the contour of a specific cross-section of the environment. Therefore, it is also called 2D LiDAR. Compared with three-dimensional point cloud, 2D LiDAR-based SLAM is a top-view LiDAR SLAM algorithm, which simplifies laser scanning and maps the data into two dimensions, which is similar to an image. 2D SLAM can save maps as images using image feature extraction and matching algorithms. Indoor sweeping robots, service robots and automatic guided vehicles mainly use SLAM based on 2D LiDAR.
In 2002, Montemerlo proposed the FastSLAM algorithm, which uses particle filters and Kalman filters to estimate robot posture and position landmarks respectively. The GMapping algorithm framework is based on the RBPF (Rao-Blackwellized Particle Filter) algorithm, which first locates and then constructs a map. Konolige developed the Karto SLAM algorithm in 2010, which is the first open source graph optimization algorithm. To address sparse decoupling, it employs height-directional optimization and non-iterative square root decomposition. However, this algorithm must build a local submap in advance in the loop closure detection part. The framework of the Hector SLAM algorithm is based on Gauss-Newton, which does not rely on odometry and is suitable for aerial or uneven road conditions. However, when the robot turns quickly, matching errors are prone to occur, and there is no loop closure detection module. In 2016, Google launched Cartographer, a sensor-equipped backpack that can generate 2D grid maps with a resolution of r = 5 cm indoors in real time. The algorithm completes front-end matching by combining relevant scan frames with gradient optimization, and uses depth-first branching and boundary search algorithms to calculate loop closure detection. Macenski built a 2D SLAM toolset called SLAM Toolbox, which provides a set of tools and capabilities for synchronous and asynchronous map modes, localization, multi-session mapping, graph optimization, reducing compute time, and prototyping lifetime and distributed mapping applications. . Table 1 summarizes the global research status of 2D LiDAR-based SLAM systems.
SLAM system based on 3D LiDAR
2D LiDAR can only scan obstacle information on the same plane, that is, the outline of a cross-section of the environment, so the information extracted from the scan is extremely limited. Multi-line LiDAR, also known as 3D LiDAR, can scan the contours of multiple cross-sections by cooperating with a rotating mechanism to simultaneously emit multiple laser beams in the vertical direction. Due to its ability to provide rich point cloud information about the surrounding environment, 3D LiDAR-based SLAM is widely used in the fields of outdoor mobile robots and autonomous driving. According to different frameworks, the 3D LiDAR-based SLAM field can be further divided into two different solutions: filter-based and graph optimization-based. Table 2 summarizes the research status of global 3D LiDAR-based SLAM systems.
Filter-based SLAM system
LINS is a lightweight LiDAR-inertial state estimator for real-time self-motion estimation. By tightly coupling a 6-axis IMU and a 3D LiDAR, it enables ground-based detection in challenging environments such as featureless scenes. The vehicle navigates robustly and efficiently. An iterative error state Kalman filter (iESKF) was developed in this system to repeatedly correct the approximate state by generating new feature correspondences in each iteration while maintaining the computational accessibility of the system. The University of Hong Kong proposed FAST-LIO2, which is based on the second generation FAST-LIO, which is also based on iESKF. In terms of filter design, FAST-LIO2 is equivalent to LINS, but the calculation of the Kalman gain is different. Shortly thereafter Gao developed Faster-LIO based on FAST-LIO2. The advantage of this algorithm over FAST-LIO2 is that it achieves greater algorithm efficiency while maintaining accuracy. This is mainly attributed to the use of iVox (incremental voxel) data structure to maintain local maps, which can effectively reduce point cloud registration time without affecting odometry accuracy. The work is currently compatible with mechanical and solid-state LiDAR. Gilmar proposed EKF-LOAM, an enhanced 3D LiDAR-based SLAM strategy that integrates wheel mileage and IMU into the SLAM process. Yuan provides VoxelMap, an efficient and probabilistically adaptive voxel mapping method for LiDAR odometry. The map consists of voxels, each containing a single planar feature, allowing for a probabilistic representation of the environment and accurate registration of new LiDAR scans. Point-LIO is a robust and high-bandwidth LiDAR inertial odometry capable of estimating extremely aggressive robot motion. Shi proposed Inv-LIO1, a machine-centered invariant EKF LiDAR-inertial odometry. The system directly integrates LiDAR and IMU measurements using an invariant observer design and Lie group theory.
SLAM system based on graph optimization
Shan introduced the LeGO-LOAM algorithm for real-time six-degree-of-freedom ground vehicle pose estimation, and Jens built a surfel-based map and estimated the robot's pose by exploiting the projection data correlation between the current scan and the surfel map's rendering model. Posture changes. Koide describes a portable human behavior measurement system based on 3D LiDAR. The system estimates the attitude of the sensor while estimating the position of the target person in the three-dimensional environment map. Shan proposed LIO-SAM, which is based on LeGO-LOAM by tightly coupling LiDAR and IMU. LIO-SAM only uses a sliding window to optimize the bias of the IMU. It then uses an additional backend to put the IMU pre-integration factor, LiDAR odometry factor, GPS factor and loop detection factor into a factor graph optimization model for joint optimization to obtain a globally consistent pose of the robot. In visual SLAM, local bundle adjustment (BA) on a sliding window of local keyframes is often used to reduce drift. Therefore, Liu formulated LiDAR BA as the process of minimizing the distance of a feature point to its matching edge or plane. This approach can significantly reduce the optimization scale and allow the use of large-scale dense planar and edge features. Wang proposed F-LOAM, a non-iterative two-stage distortion compensation method to reduce computational costs and provide a computationally efficient and accurate framework for LiDAR-based SLAM. Guo proposed E-LOAM (LOAM with Expanded Local Structural Information), which adds local point cloud information around pre-extracted geometric feature points. Wang introduced D-LIOM, a tightly coupled direct LiDAR-inertial odometry and mapping architecture. D-LIOM immediately registers scan frames to probabilistic submaps and integrates LiDAR odometry, IMU pre-integration and gravity constraints to generate local factor maps within the submap time window for real-time high-precision attitude estimation. ARTSLAM (Accurate Real-Time LiDAR SLAM) is a modular, fast and accurate LiDAR SLAM system for batch processing and online estimation. Using a three-stage algorithm, the system is able to efficiently detect and close loops in trajectories. Andrzej presented LOCUS 2.0, a robust and computationally efficient LiDAR odometry system for real-time subsurface 3D mapping.
Rotary driven LiDAR-based SLAM system
Mechanical LiDAR has a horizontal field of view (FOV, field of view) of 360°, but the vertical FOV is limited, and solid-state LiDAR has a wide vertical field of view, but a smaller horizontal field of view. Existing research usually uses multi-LiDAR solutions or rotation-driven LiDAR methods to enhance LiDAR's field of view, but the cost of the former is too high, so the latter is the mainstream trend. The rotation-driven LiDAR-based SLAM system is mainly used for applications that require panoramic scanning coverage, including surveying and mapping, underground detection, etc.
Zhen proposed a unified mapping framework (UMF) that supports multiple LiDAR types, including (1) fixed 3D LiDAR and (2) rotating 3D/2D LiDAR. The positioning module utilizes the error state Kalman filter (ESKF). and Gaussian Particle Filter (GPF) to estimate the robot state within the prior map. Mojtaba proposed LoLa-SLAM, a low-latency LiDAR SLAM framework based on LiDAR scan segmentation and concurrent matching. This framework uses segmented point cloud data from rotation-driven LiDAR in concurrent multi-threaded matching to estimate 6D pose with high update rate and low latency. Chen developed R-LIO (rotating LiDAR inertial odometry), a new SLAM algorithm that integrates rotationally driven 3D LiDAR with an IMU. R-LIO is capable of high-precision, real-time position estimation and map construction. Milad introduced Wildcat, a flexible and robust online 3D LiDAR-based SLAM system. The core of Wildcat uses continuous-time trajectory representation and an efficient pose graph optimization module to support single and multi-agent scenarios. Chanoh proposed a novel map-centric SLAM framework based on rotation-driven 3D LiDAR. With the advantages of a map-centric approach, the approach demonstrates new properties that overcome the shortcomings of existing systems related to multi-modal sensor fusion and LiDAR motion distortion. Wang proposed the online multi-calibration inertial odometer (OMC-SLIO) for SLiDAR (rotating LiDAR), which estimates multiple external parameters of LiDAR, rotating mechanism, IMU and odometry status online. Yan introduced Spin-LOAM, a tightly coupled rotationally driven 3D LiDAR-based SLAM algorithm. It uses an adaptive scan accumulation method to analyze the spatial distribution of feature points to improve the accuracy and reliability of matching. Chen demonstrated the flight-driven over-under-actuated LiDAR sensing flying robot (PULSAR), a spinning, flexible unmanned aerial vehicle (UAV) whose three-dimensional position is entirely controlled by activating a motor to generate the required thrust and torque. Table 3 summarizes the international research status of rotation-driven LiDAR-based SLAM systems.
challenge
Map quality will directly impact subsequent high-level tasks such as decision-making and planning. Among them, LiDAR-based SLAM used for mapping is a mature technology that has been thoroughly studied. Although LiDAR-based SLAM related work has made significant progress in the past few decades, there are still many challenges and problems that need to be solved. Reviewing the recent progress of LiDAR-based SLAM in robotic mapping, several aspects of challenges and future research directions in robotic mapping will be discussed below.
LiDAR-based SLAM in degraded environments. Tunnels, bridges and corridors are typical degraded environments where there are no geometric elements, the same environment and symmetrical structures. LiDAR-based SLAM systems are unable to estimate full robot 6-DOF motion in degraded environments. LiDAR-based SLAM in degraded environments represents a major challenge.
LiDAR-based lifelong SLAM in dynamic environments. All LiDAR-based SLAM techniques operate under the assumption that the environment is static. In dynamic environmental conditions, such as when a robot is creating a map, there are objects somewhere. When the robot uses the previous map for positioning, the object disappears, resulting in the failure of autonomous positioning. Lifelong mapping using LiDAR solves the problem of mapping in dynamic environments.
LiDAR-based SLAM in large-scale environments. Faced with the need for large-scale environmental mapping, using LiDAR-based SLAM solutions for multi-robot collaborative mapping can solve the problems of computational burden, global error accumulation, and risk concentration faced by single-robot SLAM.
Multi-source fusion enhances LiDAR-based SLAM. The multi-source fusion SLAM system based on 3D LiDAR is another research hotspot. Considering the inaccuracy and fragility of single sensors such as LiDAR, cameras and IMUs, researchers are increasingly developing multi-source fusion SLAM solutions.
Deep learning enhanced LiDAR-based semantic SLAM. Extensive research on deep learning enhancement of LiDAR-based SLAM systems has been conducted. The combination of deep learning and LiDAR-based SLAM in robot mapping will also be a potential research trend in the future. LiDAR-based SLAM assisted by advanced semantic information has become a basic tool in robot mapping.
Summarize
This article focuses on the research status of LiDAR-based SLAM in robotic mapping from the perspective of various LiDAR types and configurations. First, the origin of SLAM is reviewed from a historical perspective, and by comparing and analyzing the characteristics of classic SLAM based on filtering methods, a modern SLAM system framework based on optimization methods is proposed. This is followed by an extensive literature review of LiDAR-based SLAM systems in three different LiDAR forms and configurations. Compared with 3D point cloud, 2D LiDAR-based SLAM is a top-view LiDAR SLAM method. Most indoor sweeping robots, service robots and AGVs use 2D LiDAR-based SLAM. Multi-line LiDAR, also known as 3D LiDAR, enables you to scan the contours of multiple cross-sections simultaneously. Today, 3D LiDAR-based SLAM is widely used in outdoor mobile robots and autonomous driving fields. Generally speaking, rotation-driven LiDAR-based SLAM systems are mainly used for tasks that require wide-angle scanning coverage, such as surveying and mapping, underground exploration, etc.
resource
Autonomous driving and positioning related sharing
Moving object detection based on optical flow in autonomous driving
Camera extrinsic parameter calibration based on semantic segmentation
A review of autonomous vehicle positioning methods in high-speed scenarios
Patchwork++: A fast and robust ground segmentation method based on point clouds
PaGO-LOAM: Ground-based optimized lidar odometry
Multimodal road edge detection and filtering method
A framework for simultaneous calibration, positioning and mapping of multiple lidars
Extraction, mapping and long-term positioning of rods in dynamic urban environments
Motion distortion correction of non-repetitive scanning lidar
Fast and tightly coupled sparse direct radar-inertial-visual odometry
3D vehicle detection based on cameras and low-resolution lidar
Annotation tools and city datasets for 3D point cloud semantic segmentation
Basic introduction to getting started with ROS2
Automatic calibration of solid-state lidar and camera systems
Sensor fusion positioning solution of lidar+GPS+IMU+wheel speedometer
Mapping and positioning of road scenes based on sparse semantic visual features
Annotation tools and city datasets for 3D point cloud semantic segmentation
More articles can be viewed: A summary of historical articles on point cloud learning
SLAM and AR related sharing
Introduction to TOF camera principles
Introduction to TOF time-of-flight depth camera
Open source and optimized F-LOAM solution: based on optimized SC-F-LOAM
[Paper Quick Reading] AVP-SLAM: Semantic SLAM in automatic parking systems
[Quick Reading of Point Cloud Papers] StructSLAM: Structured Line Feature SLAM
Commonly used 3D depth cameras
Review and evaluation of monocular visual inertial navigation SLAM algorithm for AR equipment
SLAM Overview (4) Laser and Vision Fusion SLAM
Semantic SLAM system for real-time reconstruction by Kimera
Easy-to-extend SLAM framework-OpenVSLAM
Introduction to SLAM method based on fisheye camera
Scan QR code
Follow us
Let’s share and learn together! We look forward to friends who have ideas and are willing to share to join Knowledge Planet and inject fresh vitality into sharing. Topics shared include but are not limited to three-dimensional vision, point clouds, high-precision maps, autonomous driving, robots and other related fields.
Sharing and cooperation: WeChat "cloudpoint9527" (remarks: name + school/company + research direction) Contact email: [email protected].