PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards High-Performance Solid-State-LiDAR-Inertial Odometry and Mapping

Kailai Li, Meng Li, Uwe D. Hanebeck

arXiv:2010.13150v3cs.RO

TL;DR

LiDAR-only odometry can lose robustness under fast motion, while solid-state LiDAR adds irregular scan patterns and limited fields of view; tightly coupled fusion is also difficult to run in real time. LiLi-OM combines lightweight feature-based LiDAR odometry, hierarchical keyframe sliding-window optimization, and Horizon-specific feature extraction. The system is reported as real-time capable with superior tracking and mapping accuracy across mechanical and solid-state LiDAR experiments, while remaining a cost-effective solution; its future scope includes more scalable map representations and testing under aggressive six-DoF motion.

  • Problem

    LiDAR-only odometry is vulnerable to motion distortion, and tightly coupled fusion with solid-state LiDAR remains insufficiently investigated because irregular scans challenge conventional feature extraction.

  • Method

    LiLi-OM combines lightweight feature-based LiDAR odometry for motion estimates and keyframe selection with hierarchical keyframe sliding-window optimization that directly fuses LiDAR and IMU measurements.

  • Results

    LiLi-OM is reported as real-time capable and achieves superior tracking and mapping accuracy over state-of-the-art LiDAR and LiDAR-inertial systems for mechanical and solid-state LiDARs.

  • Takeaways & Limitations

    LiLi-OM provides a cost-effective high-performance LiDAR-inertial odometry and mapping solution using solid-state LiDAR.

  • Takeaways & Limitations

    The authors identify future work in testing aggressive six-DoF motion and improving memory and runtime efficiency for large-scale mapping with limited computational resources.

Abstract

from arXiv · show

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

LiDAR-inertial odometry addresses motion distortion and robustness limits in LiDAR-only systems, while solid-state LiDAR introduces irregular scans and narrower fields of view. LiLi-OM responds with tightly coupled fusion and a Horizon-specific feature extractor, targeting real-time, accurate mapping for both LiDAR types.

  • LiDAR provides lighting-invariant, accurate, long-range 3D perception for odometry and mapping in mobile agents.
  • Fast motion and complex scenes can degrade LiDAR-only odometry because low-rate scans are distorted by sensor motion.IMUs measure motion at higher frequency and can bridge consecutive LiDAR frames.
  • Tightly coupled fusion jointly optimizes LiDAR and IMU measurements, but exploiting every frame makes real-time processing difficult.
  • Solid-state LiDARs offer affordability and higher uniform resolution, yet commonly use irregular scans and limited fields of view.These patterns make conventional spinning-LiDAR feature extraction unsuitable and can hinder odometry in some circumstances.
  • Livox Horizon provides an 81.7° × 25.1° field of view at 10 Hz with affordability and coverage comparable to typical 64-line mechanical LiDARs.
  • LiLi-OM proposes tightly coupled fusion, Horizon-specific feature extraction, and hierarchical keyframe sliding-window optimization for real-time odometry and mapping.The authors report applicability to conventional and solid-state LiDARs, superior odometry accuracy, and released code and data.

2. System Pipeline

LiLi-OM combines a lightweight LiDAR frontend with a hierarchical keyframe-based backend that directly fuses LiDAR and IMU measurements. Sliding-window marginalization, pose-graph maintenance, and loop closure support real-time egomotion estimation and globally consistent mapping.

  • Raw scans are downsampled and rotationally de-skewed using gyroscope data before plane and edge features are extracted.
  • LiDAR and IMU measurements are directly fused in a unified keyframe-based sliding-window optimization.The fusion window usually contains several keyframes, such as three, and keyframe states include position, velocity, orientation, and IMU bias.
  • After marginalization, optimized keyframes constrain intermediate regular-frame poses, while a global pose graph incorporates all LiDAR-frame poses.
  • ICP-based keyframe loop closure can trigger global graph optimization to maintain map-wide consistency.

3. Feature-Based Solid-State LiDAR Scan-Matching

Livox Horizon’s non-repetitive brushing scan pattern prevents direct use of conventional spinning-LiDAR feature extraction. LiLi-OM therefore extracts plane and edge features using time-domain patches and covariance-based geometric tests for scan matching.

  • Scan pattern: Livox Horizon uses six vertically aligned laser diodes that sweep non-repetitively, producing an irregular brushing pattern over an 81°-wide field of view.During 100 ms frames, angular resolution reaches 0.2° to 0.4° horizontally and vertically.
  • Scan pattern: The Horizon scan covers its field of view uniformly, but its unregulated sweep differs from Mid-40’s circular, regulated coverage.
  • Feature extraction: The proposed feature extractor unfolds each sweep in time and assigns non-overlapping 6 × 7-point patches sequentially.
  • Feature extraction: Valid patch points undergo covariance eigendecomposition; patches satisfying λ1/λ2 < 0.3 are classified as plane features.
  • Feature extraction: For non-plane patches, the method searches each scan line for the point with largest curvature to extract edge features.
  • Scan matching: The time-domain features support both frontend registration and backend sliding-window fusion through point-to-edge and point-to-plane scan-matching metrics.

B A C

The method combines time-domain feature extraction with lightweight frame-to-model scan matching and a metric weighting function for LiDAR registration.

  • Point transformation: The current scan point is transformed into global coordinates using its estimated rotation and translation.The transformation is pw = R(q) pl + t, where q and t represent the current LiDAR orientation and position.
  • Plane correspondence: Plane correspondences use the nearest five plane features to fit a normal through QR decomposition before establishing the point-to-plane metric.The fitted normal is normalized from the solution of the overdetermined plane-fitting equation.
  • Metric weighting: The proposed weighting function accounts for both geometric and appearance consistency in feature associations.It uses feature-correspondence geometry together with reflectance values, with λ = 15 set from experience.
  • Livox Horizon feature extraction: Livox Horizon feature extraction operates on time-domain sweep patches and outputs edge and plane feature sets.Algorithm 1 takes a single sweep as input and produces edge and plane features; its covariance eigendecomposition checks plane and edge candidates.
  • Feature-based scan-matching: The frontend estimates motion by minimizing point-to-edge and point-to-plane metrics against a local feature map.The map is updated from typically 20 recent frames, while optimization time is restricted for fast motion estimation.

4. Tightly-Coupled LiDAR-Inertial Fusion via Keyframe-Based Sliding Window Optimization

The backend directly fuses LiDAR and pre-integrated IMU measurements over keyframe-based sliding windows, using marginalization to control computation while retaining prior information.

  • Keyframe-based fusion hierarchy: Keyframes establish sliding windows where LiDAR and pre-integrated IMU measurements are fused jointly through nonlinear optimization.Regular-frame poses are initialized from IMU measurements, while older keyframe poses provide constraints as the window advances.
  • Keyframe-based fusion hierarchy: Keyframes are selected when feature overlap with the local map falls below 60% or the interval since the last keyframe exceeds a set threshold.Limiting the interval between keyframes helps mitigate drift in IMU pre-integration.
  • Sliding window optimization: Marginalization removes the oldest keyframe through a Schur complement and adds a new prior factor to carry its information into the next window.This bounds computational burden without substantial information loss.
  • Measurement fusion: The LiDAR residual uses point-to-edge and point-to-plane geometric constraints with fixed feature correspondences and normalized feature-type weights.The local map is updated from optimized poses as the window slides.
  • Measurement fusion: The IMU residual constrains relative motion between consecutive keyframes using raw inertial readings pre-integrated over each keyframe interval.Pre-integration avoids repropagating inertial states whenever the optimization window slides.
  • Global mapping: After window optimization, regular LiDAR-frame poses are refined locally, and resulting poses enter a global pose graph while keyframe feature maps support mapping.Loop-closure detection searches the global graph and can invoke global graph optimization.

5. Evaluation

The evaluation covers conventional and solid-state LiDARs across public and newly recorded data sets, with LiLi-OM achieving real-time operation and strong tracking, mapping, and drift performance.

  • 5.2. Public data set: LiLi-OM⋆ achieves the best tracking accuracy for all evaluated public-data sequences in real time.Comparisons include LiDAR odometry and LiDAR-inertial systems on UTBM, UrbanLoco, and UrbanNav data sets.
  • 5.3.1. FR-IOSB data set: The Livox-Xsens low-cost setup performs equally well as the high-end setup, while LiLi-OM and LiLi-OM⋆ deliver superior tracking accuracy on FR-IOSB sequences.The sequences cover structured scenes, bushes, and a long trajectory; LiLi-OM⋆ also produces superior mapping quality on Long.
  • 5.3.2. KA-Urban data set: LiLi-OM without loop closure delivers much less drift than LiHo on KA-Urban sequences, while loop closure yields very small end-to-end errors.The KA-Urban evaluation uses five Karlsruhe sequences with locations registered from satellite images.
  • 5.3.2. KA-Urban data set: Removing IMU constraints remains superior to pure LiDAR LiHo but is inferior to standard LiDAR-inertial fusion on long-distance sequences.The comparison evaluates configurations with loop closure both on and off.
  • 5.3.2. KA-Urban data set: 496 m × 312 m is mapped accurately with global consistency on Schloss-2, and Sequence East retains accurate odometry and mapping under fast, dynamic motion.Both results use the proposed low-cost Livox-Xsens sensor suite.
  • 5.4. Runtime: LiLi-OM runs in real time at the LiDAR frame rate on all evaluated data sets using a four-core Intel Core i5-7300HQ laptop; backend fusion dominates runtime.Preprocessing and frontend scan registration are described as lightweight.

6. Conclusion

The paper presents a real-time LiDAR-inertial fusion system for conventional and solid-state LiDARs, including a Livox Horizon feature extractor and hierarchical sliding-window optimization. Evaluations report superior tracking and mapping accuracy over state-of-the-art systems, while future work targets broader deployment and more efficient large-scale mapping.

  • 6. Conclusion: The system directly fuses LiDAR and pre-integrated IMU measurements through a keyframe-based hierarchical sliding-window scheme.Regular-frame poses are obtained through factor graph optimization after keyframe optimization.
  • 6. Conclusion: LiLi-OM is designed for conventional LiDARs and small-FoV solid-state LiDARs, with feature extraction tailored to Livox Horizon’s irregular scan pattern.The evaluations include public conventional-LiDAR data sets and Livox Horizon experiments.
  • 6. Conclusion: The proposed system is real-time capable and delivers superior tracking and mapping accuracy over state-of-the-art LiDAR and LiDAR-inertial odometry systems.LiLi-OM is presented as a cost-effective solid-state-LiDAR solution.
  • 6. Conclusion: Future work includes testing aggressive six-DoF motion on unmanned aerial vehicles and improving memory and runtime efficiency for large-scale mapping.Suggested map representations include TSDF volumetric mapping and geometric primitives.
Loading 2010.13150v3…