Source-linked AI summary
Towards High-Performance Solid-State-LiDAR-Inertial Odometry and Mapping
Kailai Li, Meng Li, Uwe D. Hanebeck
TL;DR
Low-rate LiDAR scans can distort with sensor motion, and tightly coupled solid-state LiDAR–inertial fusion had not been reported. LiLi-OM combines LiDAR and IMU measurements in a keyframe-based sliding-window system and is reported to run in real time with superior tracking and mapping accuracy over state-of-the-art systems.
Problem
Low-rate LiDAR scans can distort with sensor motion, and tightly coupled solid-state LiDAR–inertial fusion had not been reported.
Source: §1 · pp. 1, 3Method
LiLi-OM uses lightweight feature-based LiDAR odometry for motion estimates and adaptive keyframe selection, then directly fuses LiDAR and IMU measurements through hierarchical keyframe-based sliding-window optimization.
Source: Contributions · p. 3Results
LiLi-OM runs in real time and delivers superior tracking and mapping accuracy over state-of-the-art LiDAR/LiDAR-inertial odometry systems.
Source: §6 · p. 15Takeaways & Limitations
LiLi-OM is presented as a cost-effective solution for high-performance LiDAR-inertial odometry and mapping using solid-state LiDAR.
Source: §6 · p. 15Large-scale mapping with limited computational resources remains an opportunity to improve memory and runtime efficiency using advanced map representations.
Source: §6 · p. 15
Abstract
from arXiv · showhide
We present a novel tightly-coupled LiDAR-inertial odometry and mapping scheme for both solid-state and mechanical LiDARs. As frontend, a feature-based lightweight LiDAR odometry provides fast motion estimates for adaptive keyframe selection. As backend, a hierarchical keyframe-based sliding window optimization is performed through marginalization for directly fusing IMU and LiDAR measurements. For the Livox Horizon, a newly released solid-state LiDAR, a novel feature extraction method is proposed to handle its irregular scan pattern during preprocessing. LiLi-OM (Livox LiDAR-inertial odometry and mapping) is real-time capable and achieves superior accuracy over state-of-the-art systems for both LiDAR types on public data sets of mechanical LiDARs and in experiments using the Livox Horizon. Source code and recorded experimental data sets are available at https://github.com/KIT-ISAS/lili-om.
1. Introduction
The introduction motivates combining LiDAR’s robust, accurate perception with higher-frequency IMU measurements to address motion distortion and the limits of LiDAR-only odometry. It presents a real-time, tightly coupled system for conventional and solid-state LiDARs, with feature extraction tailored to Livox Horizon’s irregular scans.
1. Introduction
- LiDAR offers lighting-invariant, accurate, long-range perception, but low scan rates distort point clouds during sensor motion; higher-frequency IMUs can bridge frames and improve odometry robustness and accuracy.Conventional mechanical LiDARs have limited vertical resolution, while high-resolution products can be prohibitively expensive; newer solid-state sensors are more affordable but often have irregular scans and smaller fields of view.Source: §1 · pp. 1–3
- Prior tightly coupled methods jointly optimize LiDAR and IMU data, but processing every frame can hinder real-time operation, while factor-based fusion may lose landmark constraints and measurement correlations.The cited factor-graph system also relies on high-frequency nine-axis IMU readings and may require additional modalities such as GPS for precise localization.Source: §1 · p. 2
- Livox Horizon’s irregular, non-repetitive scans make spinning-LiDAR feature extraction inapplicable, while its limited field of view can impair odometry during fast motion or with insufficient features.The paper notes that Horizon sweeps in an unregulated manner, unlike the circular Livox Mid-40 pattern addressed by prior work.Source: §1 · p. 3
Contributions
- The paper proposes a real-time, tightly coupled LiDAR-inertial odometry and mapping system for conventional and solid-state LiDARs, with hierarchical keyframe sliding-window optimization and a Livox Horizon feature extractor.The authors report superior odometry accuracy over existing systems and release source code and new Livox Horizon–Xsens MTi-670 datasets.Source: Contributions · p. 3
2. System Pipeline
The system estimates LiDAR motion through lightweight scan matching while fusing LiDAR and IMU measurements in a keyframe-based sliding-window backend. It also maintains global consistency through keyframe loop closure and pose-graph optimization.
System Pipeline
- The frontend extracts plane and edge features, estimates motion with frame-to-model scan matching, then uses those estimates for translational de-skewing and adaptive keyframe selection.Preprocessing downsamples scans and uses gyroscope data to correct rotational distortion.Source: §2 · p. 3
- The backend fuses LiDAR and IMU measurements through keyframe-based sliding-window optimization, typically covering several keyframes.As the window slides, the two oldest optimized keyframes constrain in-between regular-frame poses through factor-graph optimization; preintegrated inertial measurements initialize poses.Source: §2 · p. 4
- Each optimized keyframe state comprises position, orientation, velocity, and accelerometer and gyroscope biases.Orientation is represented by a unit quaternion.Source: §2 · p. 4
- The system maintains a global pose graph, checks for loop closure by keyframe-based ICP, and invokes global graph optimization when needed for a globally consistent map.Source: §2 · p. 4
3. Feature-Based Solid-State LiDAR Scan-Matching
LiLi-OM adapts feature-based scan matching to the Livox Horizon’s irregular sweep pattern and uses edge and plane residuals for LiDAR motion estimation. Its frontend limits optimization time for fast estimates, while the backend pursues accuracy through sensor fusion.
3.1. Feature extraction for irregular scan pattern
- Horizon’s six vertically aligned lasers produce an irregular, non-repetitive brushing scan, with 0.2°–0.4° angular resolution over 100 ms and FoV coverage similar to 64-line spinning LiDARs.Its unregulated sweep contrasts with the Mid-40’s regulated circular FoV scan.Source: §3.1 · p. 4
- The method divides each sweep into non-overlapping 6 × 7-point time-domain patches, classifies planar patches by covariance eigenvalues, and extracts edge points from scan lines in non-planar patches.Plane features are selected when λ1/λ2 < 0.3; edge features are selected when λ2/λ3 < 0.25, otherwise no feature is extracted.Source: §3.1 · pp. 4–5
3.2. Point-to-edge and point-to-plane metric
- Both frontend registration and backend sliding-window fusion use scan matching to incorporate LiDAR measurements for egomotion estimation.Source: §3.2 · p. 5
Point-to-edge metric
- For edge matching, five nearest map-edge points are fitted by eigendecomposition, and two points on the fitted line define the point-to-edge metric.The current transformed scan point is evaluated against the fitted line; the passage gives δ = 0.1 as the typical offset.Source: Point-to-edge metric · p. 6
Point-to-plane metric
- For plane matching, five nearest map-plane points define a fitted plane via an overdetermined linear system solved by QR decomposition, yielding a normalized plane normal for the point-to-plane metric.The scan point is transformed into world coordinates before applying the metric.Source: Point-to-plane metric · p. 6
3.3. Metric weighting function
- The proposed residual weighting function evaluates feature-association quality using geometric direction or normal consistency and reflectance consistency, with λ = 15 set from experience.It distinguishes edge and plane correspondences and compares their local feature geometry and reflectance values.Source: §3.3 · pp. 6–7
3.4. Feature-based scan-matching
- The frontend estimates pose by trust-region frame-to-model matching against a local feature map, limiting optimization time to correct translational scan distortion and select keyframes.The map typically contains 20 recent frames; the backend fusion scheme is used to pursue state-estimation accuracy.Source: §3.4 · p. 7
4. Tightly-Coupled LiDAR-Inertial Fusion via Keyframe-Based Sliding Window Optimization
The backend directly fuses LiDAR and pre-integrated IMU measurements in keyframe-based sliding windows, using marginalization to control computation and local optimization to recover regular-frame poses.
4.1. Keyframe-based fusion hierarchy
- Keyframes form sliding windows for unified LiDAR–IMU optimization; local factor graphs use the two oldest keyframe poses and IMU-initialized regular poses to recover poses at LiDAR frequency.A frame becomes a keyframe when feature-map overlap falls below 60% or the interval since the previous keyframe exceeds a set limit (e.g., two regular frames), helping limit IMU pre-integration drift.Source: §4.1 · pp. 8–9, Figure 6 · p. 9
4.2. Sliding window optimization for keyframes
- Keyframe states are estimated by MAP optimization over a τw-keyframe window, directly fusing keyframe LiDAR and pre-integrated IMU measurements with prior, LiDAR, and IMU residuals.Source: §4.2 · p. 9
Prior factor
- The oldest keyframe and its measurements are marginalized via Schur complement, and a new prior carries its estimate into the next window.Source: Prior factor · p. 9
LiDAR term
- The LiDAR term adds geometric constraints using fixed feature correspondences and point-to-edge or point-to-plane metrics, then updates the local map from optimized poses.The described alignment uses edge observations from recent 30 keyframes and normalized correspondence weights.Source: LiDAR term · pp. 9–10
IMU term
- The IMU term constrains relative motion with pre-integrated readings between consecutive keyframes, avoiding repeated state propagation as the window slides.Optimized keyframes constrain local graph optimization of regular frames; resulting poses enter a global pose graph, where ICP-verified loop closures trigger global optimization and update the backend local map.Source: IMU term · p. 10
5. Evaluation
LiLi-OM and its spinning-LiDAR variant are evaluated on public datasets and real-world Livox Horizon experiments. The section reports tracking, mapping, and runtime results, including real-time performance across evaluated devices.
5.1. Implementation and evaluation setup
- The system is implemented in ROS with Ceres optimization and GTSAM loop-closure pose-graph rectification; LiLi-OM⋆ uses spinning-LiDAR preprocessing from.Evaluations use public conventional-LiDAR datasets and Livox Horizon experiments, with all LiDAR frame rates at 10 Hz.Source: §5.1 · p. 10
5.2. Public data set
- LiLi-OM⋆ achieves the best tracking accuracy on every public-dataset sequence in real time, while competing systems show the stated limitations.The evaluation uses final-trajectory APE RMSE against ground truth; LIO-SAM is inapplicable to UTBM because it requires nine-axis IMU readings, and LIOM fails on UrbanNav and cannot run in real time with recommended configurations.Source: §5.2 · p. 11, Table 1 · p. 11
5.3. Experiment
- The Livox Horizon–Xsens suite costs about 1700 Euros (Q1, 2020) and is deployed on a backpack platform for large-scale urban testing.Five Karlsruhe sequences use end-to-end locations registered from satellite images.Source: §5.3 · p. 11, §5.3.2 · p. 12
5.3.1. FR-IOSB data set
- On FR-IOSB, both LiLi-OM versions deliver superior tracking accuracy, with the low-cost Livox-Xsens setup performing as well as the high-end setup in the same scenario.On sequence Long, LiLi-OM⋆ also produces superior mapping quality through its sensor-fusion scheme.Source: §5.3.1 · p. 11, Table 2 · p. 12, Figure 8 · p. 12
5.3.2. KA-Urban data set
- LiLi-OM without loop closure drifts less than LiHo; loop closure yields very small end-to-end errors, and removing IMU constraints performs between LiHo and standard LiLi-OM.The IMU-removed configuration is evaluated on the three long-distance sequences with loop closure on and off.Source: §5.3.2 · pp. 12–13, Table 3 · p. 13
- LiLi-OM accurately maps a 496 m × 312 m area with global consistency, and produces accurate odometry and maps during fast, dynamic cycling in eastern Karlsruhe.The reported mapping examples include Schloss-2 and the East sequence.Source: §5.3.2 · p. 13, Figure 9 · p. 13, Figure 10 · p. 14
5.4. Runtime
- LiLi-OM runs in real time at LiDAR frame rate across devices on a four-core laptop, with backend fusion dominating runtime.Backend LiDAR constraints are formed from voxel-grid-downsampled features; Livox Horizon’s relatively uniform field of view yields more plane features than edges on average.Source: §5.4 · pp. 13–14, Table 4 · p. 14, Table 5 · p. 14
6. Conclusion
LiLi-OM is a real-time LiDAR-inertial odometry and mapping system that directly fuses LiDAR and pre-integrated IMU measurements in hierarchical sliding-window optimization. Evaluations report superior tracking and mapping accuracy over state-of-the-art systems across conventional-LiDAR datasets and Livox Horizon experiments.
6. Conclusion
- A keyframe-based hierarchical scheme directly fuses LiDAR and pre-integrated IMU measurements through sliding-window optimization, with regular-frame poses obtained by factor graph optimization.The system is proposed for real-time LiDAR-inertial odometry and mapping.Source: §6 · p. 14
- The system applies to conventional and small-FoV solid-state LiDARs, with a novel feature extractor designed for the Livox Horizon’s irregular scan pattern.The conclusion describes the Horizon as a newly released, spinning-free solid-state LiDAR with a much lower price than conventional 3D LiDARs.Source: §6 · p. 14
- LiLi-OM is real-time capable and delivers superior tracking and mapping accuracy over state-of-the-art LiDAR/LiDAR-inertial odometry systems.The evaluations use public datasets from conventional LiDARs and experiments with the Livox Horizon.Source: §6 · pp. 14–15
- Future work includes testing the lightweight Livox-Xsens suite on UAVs during aggressive six-DoF motion and using advanced map representations to improve memory and runtime efficiency.Suggested mapping options include TSDF volumetric mapping and geometric primitives, especially in man-made environments.Source: §6 · p. 15