Source-linked AI summary
F-LOAM: Fast LiDAR Odometry And Mapping
Han Wang, Chen Wang, Chun-Lin Chen, Lihua Xie
TL;DR
LiDAR SLAM must balance localization accuracy with computational efficiency, while existing iterative scan matching and map refinement are costly. F-LOAM uses non-iterative two-stage distortion compensation with edge and planar feature matching, and reports competitive accuracy with processing above 10 Hz.
Problem
LiDAR SLAM requires both accurate localization and computational efficiency, but existing scan-to-scan and scan-to-map modules rely on computationally expensive iterative calculation.
Method
F-LOAM uses non-iterative two-stage distortion compensation, extracts edge and planar features, and incorporates local smoothness into iterative pose optimization.
Results
F-LOAM achieves competitive localization accuracy at low computational cost, including 0.80% average translational error and 0.0048 deg /m average rotational error over 11 sequences.
Takeaways & Limitations
F-LOAM provides a practical real-time LiDAR SLAM solution with a reported trade-off between localization performance and processing speed.
Abstract
from arXiv · showhide
Simultaneous Localization and Mapping (SLAM) has wide robotic applications such as autonomous driving and unmanned aerial vehicles. Both computational efficiency and localization accuracy are of great importance towards a good SLAM system. Existing works on LiDAR based SLAM often formulate the problem as two modules: scan-to-scan match and scan-to-map refinement. Both modules are solved by iterative calculation which are computationally expensive. In this paper, we propose a general solution that aims to provide a computationally efficient and accurate framework for LiDAR based SLAM. Specifically, we adopt a non-iterative two-stage distortion compensation method to reduce the computational cost. For each scan input, the edge and planar features are extracted and matched to a local edge map and a local plane map separately, where the local smoothness is also considered for iterative pose optimization. Thorough experiments are performed to evaluate its performance in challenging scenarios, including localization for a warehouse Automated Guided Vehicle (AGV) and a public dataset on autonomous driving. The proposed method achieves a competitive localization accuracy with a processing rate of more than 10 Hz in the public dataset evaluation, which provides a good trade-off between performance and computational cost for practical applications.
I. INTRODUCTION
LiDAR SLAM is valuable for robotics but remains constrained by environmental robustness and computational cost. F-LOAM addresses these practical concerns with lightweight distortion compensation and feature-based pose estimation, achieving competitive accuracy and speed.
- LiDAR SLAM supports applications including autonomous driving, drone inspection, and warehouse manipulation because it is accurate and robust to illumination and weather changes.
- Existing LiDAR SLAM systems remain limited in practical applications by robustness across indoor, outdoor, static, and dynamic environments.One cited example reports good indoor performance but substantially lower localization accuracy outdoors.
- Computational cost is another challenge because resource-limited robotic platforms must perform high-frequency localization and path planning concurrently.
- F-LOAM combines feature extraction, distortion compensation, pose optimization, and mapping in a lightweight LiDAR SLAM framework for practical real-time use.Its non-iterative two-stage distortion compensation replaces computationally inefficient iterative distortion compensation.
- F-LOAM considers local feature geometry during iterative pose estimation and achieves competitive localization accuracy at low computational cost across indoor and outdoor experiments.The paper describes the method as one of the most accurate and fastest open-sourced methods in the KITTI benchmark.
II. RELATED WORK
LiDAR SLAM research centers on point-cloud matching, with feature-based methods improving efficiency over raw matching. Extensions add loop closure, sensor fusion, or learned features, while F-LOAM presents its overall processing framework.
- Raw point-cloud methods such as ICP iteratively minimize correspondence distances, but IMLS-SLAM requires 1.2 s per frame and is far from real-time performance.
- LOAM established feature-based point-to-surface and point-to-edge matching using locally smoothness-based edge and surface extraction.Subsequent methods extend this feature-based formulation to improve performance.
- Loop closure detects repetitive places and corrects accumulated odometry drift in large-scale SLAM scenarios.
- IMU and GPS fusion can improve localization accuracy over LOAM outdoors, but multi-sensor systems require synchronization and comprehensive calibration.
- Deep-learning methods replace handcrafted feature selection with learned point-cloud features, and CAE-LO reports improved matching success rate while retaining feature-point matching for odometry.
- F-LOAM organizes feature extraction, distortion compensation, global feature estimation, laser odometry, and mapping into one proposed LiDAR SLAM approach.
A. Sensor Model and Feature Extraction
F-LOAM models mechanical LiDAR scans as sequential measurements from a rotating vertical beam array, then selects horizontal edge and surface features using local smoothness.
- A mechanical 3D LiDAR uses an M-reading vertical laser array that rotates horizontally at constant speed while measurements arrive sequentially.
- Raw ICP matching is sensitive to noise and dynamic objects and is computationally inefficient for scans containing tens of thousands of points.
- F-LOAM retains surface and edge features while discarding noisy or less significant points to improve matching accuracy and efficiency.
- The method focuses on horizontal-plane structure because horizontal features are more distinctive and less likely to produce false detections there.
- Local smoothness is computed from adjacent horizontal points, which can be collected by point ID rather than local searching.The experiment uses five clockwise and five counterclockwise neighboring points.
- High-smoothness points are selected as edge features, while low-smoothness points are selected as surface features.Flat walls have small smoothness values, whereas corners and edges have large values.
B. Motion Estimation and Distortion Compensation
F-LOAM replaces iterative scan-to-scan distortion correction with a two-stage procedure that predicts short-term motion, then recomputes distortion after pose estimation.
- LOAM and LeGO-LOAM iteratively estimate transformations between consecutive scans to correct distortion, which is computationally inefficient.
- F-LOAM first assumes constant angular and linear velocity over the short interval between scans to predict motion and correct distortion.The assumption is motivated by LiDAR scan rates above 10 Hz and short inter-scan elapsed time.
- After pose estimation, F-LOAM recomputes distortion and updates the resulting undistorted features in the final map.
- The two-stage distortion compensation achieves similar localization accuracy to iterative compensation at substantially lower computational cost.
C. Pose Estimation
F-LOAM estimates pose by aligning undistorted edge and planar features with separately maintained global maps. It uses local smoothness-weighted geometric residuals and iterative nonlinear optimization to refine the pose.
- Map alignment: Undistorted edge and planar features are aligned with separately maintained global edge and planar maps.Both maps are stored in 3D KD-trees, and nearby points are collected to estimate corresponding global lines and planes.
- Geometric correspondence: Global lines use the largest-eigenvalue orientation and nearby-point center, while global plane normals use the smallest-eigenvalue eigenvector.The line position is the geometric center of nearby edge points; plane position and surface normal are estimated analogously.
- Residual minimization: Pose optimization minimizes distances from feature points to corresponding global edges and planes.The edge residual uses a point-to-line distance, while the planar residual uses a point-to-plane distance.
- Weighted matching: Local smoothness weights supplement geometric distances because consistently extracted edge and planar features can improve matching.Higher-smoothness edge features and lower-smoothness planar features are emphasized through a reused smoothness-based weight function.
- Iterative optimization: The nonlinear pose problem is solved with Gauss–Newton optimization and repeated correspondence updates until convergence.The Jacobian is estimated using a left perturbation model with δξ ∈ se(3).
D. Mapping Building & Distortion Compensation Update
F-LOAM maintains separate global edge and planar maps, updates them with re-computed undistorted features, and limits map growth through keyframe selection and voxel downsampling.
- Mapping Building: The global edge and planar maps are updated only when translation or rotation exceeds predefined keyframe thresholds.Keyframe-based updates reduce computational cost compared with updating the map frame by frame.
- Distortion Compensation Update: In the second stage, distortion is re-computed using the optimized pose T∗k.The resulting undistorted edge and planar features are then inserted into their respective global maps.
- Mapping Building: After each map update, a 3D voxelized grid downsamples the maps to prevent memory overflow.
A. Experiment Setup
F-LOAM is evaluated on KITTI and warehouse-related scenarios, with comparisons covering accuracy, processing speed, and trajectory tracking against ground truth and existing methods.
- Evaluation scope: The evaluation covers a large-scale outdoor KITTI dataset and medium-scale indoor warehouse environments.KITTI contains 23,201 frames and 22 km of travel over 11 sequences.
- KITTI evaluation: 0.80% average translational error and 0.0048 deg /m average rotational error are achieved over 11 KITTI sequences.These metrics are computed using frame-pair sets comparing estimated and true LiDAR poses.
- Method comparison: F-LOAM is compared with LiDAR and visual SLAM methods using average processing time measured from feature extraction through estimated odometry output.Compared methods include LOAM, A-LOAM, HDL-Graph-SLAM, IMLS-SLAM, LeGO-LOAM, and VINS-MONO.
- Warehouse evaluation: Warehouse experiments include simulated and actual AGV settings, with trajectory comparisons against ground truth and a reported 2 cm average localization error indoors.The actual AGV experiment uses an Intel NUC mini computer, while the indoor accuracy evaluation uses VICON ground truth.
- Method comparison: F-LOAM achieves the second-highest accuracy and an average processing rate above 10 Hz among the compared methods.The evaluation characterizes this as a trade-off between computational cost and localization accuracy.
C. Experiment on warehouse logistics
The warehouse logistics experiment targets an autonomous AGV that can perform transportation tasks while localizing itself in a complex manufacturing environment.
- Application objective: The AGV is designed to replace human-labour-dominated manufacturing by carrying out daily transportation tasks.The experiment focuses on autonomous warehouse logistics.
- Application objective: The robot must actively localize itself in a complex environment to support these warehouse operations.
- Application objective: The stated application context is an autonomous warehouse robot for manufacturing workflows.
1) Simulation:
F-LOAM is tested in simulated warehouse environments containing static and dynamic objects, where it tracks a rapidly moving AGV against ground truth and is also implemented on a real AGV.
- Simulation setup: The simulation uses Gazebo and Ubuntu 18.04 with a virtual Pioneer robot and Velodyne VLP-16 LiDAR.The warehouse reconstruction includes moving human workers, shelves, and machines.
- Simulation setup: The simulated robot moves at a maximum speed of 2 m/s while F-LOAM trajectory is compared with ground truth.The trajectories are plotted in green and red, respectively.
- Simulation results: F-LOAM tracks the AGV at high speed in a dynamic warehouse environment with walking human operators.
- Real-world evaluation: The method is additionally implemented on an actual AGV in a smart-manufacturing warehouse.The environment includes an auto charging station, material handling area, and manufacturing station.
3) Performance evaluation:
The evaluation shows that F-LOAM accurately tracks indoor robot pose and reduces computational cost through its proposed distortion compensation, while achieving strong overall results across indoor and outdoor tests.
- Indoor localization: 2 cm average localization accuracy was achieved against VICON ground truth in the indoor room experiment.The F-LOAM and ground-truth trajectories are reported as accurately aligned.
- Distortion compensation: The proposed distortion compensation was much faster than LOAM with motion compensation while providing slightly better localization accuracy.The ablation compares no compensation, iterative LOAM compensation, and the proposed motion compensation.
- Overall evaluation: F-LOAM achieved an average localization accuracy of 2 cm indoors and ranked among the most accurate and fastest open-sourced methods on KITTI.The evaluation covered simulation, an indoor AGV test, and an outdoor autonomous-driving test.