PaperlayerView exact revision on arXiv

Source-linked AI summary

LiDAR Odometry Survey: Recent Advancements and Remaining Challenges

Dongjae Lee, Minwoo Jung, Wooseong Yang, Ayoung Kim

arXiv:2312.17487v1cs.RO

TL;DR

Precise odometry is important for robot navigation, particularly when GPS is unavailable, and LiDAR offers 3D sensing that is unaffected by lighting conditions. This survey organizes LiDAR odometry research across sensor modalities and synthesizes its challenges, datasets, and evaluation; its conclusion is that system choice depends on application needs and available resources.

  • Problem

    The survey addresses incomplete coverage of LiDAR odometry methods and the need to assess their challenges, datasets, and evaluation approaches.

    Source: §1 · pp. 2–3

  • Method

    The paper surveys LiDAR odometry by sensor integration approach, reviews challenges and datasets, and discusses evaluation metrics and benchmark results.

    Source: §1 · p. 3

  • Results

    The survey concludes that no one LiDAR odometry approach fits all applications, with choices involving trade-offs between resources and performance.

    Source: §9 · p. 21

  • Takeaways & Limitations

    System selection should reflect application requirements and available resources, with the paper describing LiDAR–IMU fusion as a balanced general choice.

    Source: §9 · p. 21

    The survey notes limited work integrating diverse sensor modalities and identifies asynchronous sensor data as a synchronization challenge.

    Source: §7.5 · p. 16

Abstract

from arXiv · show

Odometry is crucial for robot navigation, particularly in situations where global positioning methods like global positioning system (GPS) are unavailable. The main goal of odometry is to predict the robot's motion and accurately determine its current location. Various sensors, such as wheel encoder, inertial measurement unit (IMU), camera, radar, and Light Detection and Ranging (LiDAR), are used for odometry in robotics. LiDAR, in particular, has gained attention for its ability to provide rich three-dimensional (3D) data and immunity to light variations. This survey aims to examine advancements in LiDAR odometry thoroughly. We start by exploring LiDAR technology and then scrutinize LiDAR odometry works, categorizing them based on their sensor integration approaches. These approaches include methods relying solely on LiDAR, those combining LiDAR with IMU, strategies involving multiple LiDARs, and methods fusing LiDAR with other sensor modalities. In conclusion, we address existing challenges and outline potential future directions in LiDAR odometry. Additionally, we analyze public datasets and evaluation methods for LiDAR odometry. To our knowledge, this survey is the first comprehensive exploration of LiDAR odometry.

1 Introduction

LiDAR’s ability to measure surroundings in 3D without being affected by lighting conditions helped drive its use for odometry, as researchers sought precise location data for autonomous robots. This survey addresses the coverage gap in earlier reviews by organizing LiDAR odometry across sensor modalities and examining challenges, datasets, evaluation metrics, and benchmark results.

1 Introduction

  • Robotic odometry developed from wheel encoders and dead reckoning toward range- and visual-sensor approaches, alongside advances such as ICP scan registration.Wheel odometry accuracy was constrained by wheel slippage and algorithmic inaccuracies.Source: §1 · p. 1
  • LiDAR gained attention because it measures surrounding space in 3D while remaining unaffected by lighting, unlike visual odometry’s stated low-light limitation.The passage connects this development to the need for precise location data in autonomous robot decision-making.Source: §1 · p. 2
  • Earlier surveys addressed narrower areas, including visual odometry, UAV platforms, automotive LiDAR, LiDAR-only odometry, or indoor LiDAR SLAM, leaving other LiDAR odometry methodologies uncovered.The authors present their review as addressing this gap by covering both LiDAR-only methods and approaches integrating other sensors.Source: §1 · p. 2
  • The survey covers LiDAR-only, LiDAR–IMU, multiple-LiDAR, and other sensor-fusion approaches, alongside challenges, public datasets, evaluation metrics, and benchmarks.Its stated contributions also include reviewing LiDAR technology and outlining future research directions.Source: §1 · pp. 2–3

2 LiDAR Preliminary

The survey introduces LiDAR’s measurement of distances and generation of 3D point clouds, then classifies sensors by imaging architecture and measurement principle. It describes how mechanical, scanning solid-state, ToF, and FMCW designs differ in scanning patterns and operational trade-offs.

2 LiDAR Preliminary

  • LiDAR odometry depends on understanding sensor fundamentals and categories, which this section introduces as context for the survey.It investigates fundamental principles and different categories of LiDAR sensors.Source: §2 · p. 3

2.1 Light Detection And Ranging

  • LiDAR measures distance from returned laser pulses and can produce point clouds that map objects and surroundings in 3D.The distance calculation uses the pulse’s return time and the constant speed of light.Source: §2.1 · p. 3

2.2 LiDAR Categorization

  • LiDARs are categorized by imaging architecture—mechanical, scanning solid-state, or flash—and by measurement principle, including ToF, AMCW, and FMCW.The survey focuses on mechanical and scanning solid-state imaging, and ToF and FMCW measurement principles.Source: §2.2 · p. 3

2.2.1 Imaging Mechanisms

  • Mechanical LiDAR uses moving components, while scanning solid-state designs steer beams without mechanical rotation; Risley-prism systems offer compactness and robustness but have limited FOV.Mechanical components can degrade, slow acquisition, and increase vulnerability to vibration and shock; the Risley-prism design’s scanning pattern is illustrated alongside a mechanical sensor’s repetitive pattern.Source: §2.2.1 · pp. 3–4, Figure 2 · p. 4

2.2.2 Measurement Principles

  • ToF measures pulse return time, whereas FMCW analyzes reflected-light frequency shifts to measure distance and relative velocity.ToF can be affected by external light; FMCW is described as resilient to interference but more intricate and potentially pricier.Source: §2.2.2 · p. 4

3 LiDAR-Only Odometry

LiDAR-only odometry estimates motion from consecutive LiDAR scans using direct, feature-based, or deep-learning-based matching. The section surveys their mechanisms and limitations, including correspondence difficulties and scan-matching accuracy constraints.

3 LiDAR-Only Odometry

  • LiDAR-only odometry estimates a robot’s position from consecutive scans and is grouped by direct, feature-based, and deep-learning-based scan matching.Table 1 summarizes the LiDAR-only literature using these three matching categories.Source: §3 · p. 4, Table 1 · p. 5

3.1 Direct Matching

  • Direct matching estimates transformations between consecutive scans, but ICP can be sensitive to noise and local minima and computationally expensive.ICP iteratively minimizes an error metric, typically squared distances between matched point pairs; scan-to-scan matching accuracy is inherently limited, prompting use of scan-to-map methods as well.Source: §3.1 · pp. 4–5
  • Direct-matching variants improve registration through robust ICP or probabilistic NDT, while other methods use scan-to-map matching or continuous-time trajectories.Examples include trimmed-residual ICP, point-to-plane and Generalized-ICP, probabilistic NDT, and CT-ICP’s interpolation of points between scan-end poses.Source: §3.1 · pp. 5–6

3.2 Feature-based Matching

  • Feature-based methods use selected point-cloud features to estimate transformations, potentially improving speed and performance by reducing outliers, but depend on selecting good features.Examples include LOAM’s edge and planar features and methods using surface normals, semantic constraints, or multiple feature types.Source: §3.2 · p. 6

3.3 Deep Learning-based Matching

  • Deep-learning approaches target correspondence matching; surveyed methods predict scan geometry or uncertainty, select keypoint matches, or estimate relative pose between frames.LO-Net predicts normals and dynamic regions; LodoNet back-projects matched keypoints; VertexNet and PoseNet estimate point uncertainty and relative pose, with an uncertainty-weighted ICP loss.Source: §3.3 · pp. 6–7

4 LiDAR-Inertial Odometry

LiDAR-inertial odometry combines LiDAR with IMU measurements to estimate coarse motion and improve pose estimation, using loosely or tightly coupled fusion. The survey describes their trade-offs and the range of tightly coupled methods developed for efficiency, accuracy, and challenging environments.

4 LiDAR-Inertial Odometry

  • LiDAR-inertial odometry fuses LiDAR and IMU data either loosely or tightly, with IMU angular velocity and linear acceleration supporting coarse motion estimates and pose estimation.Tightly coupled estimation can use more constraints and potentially improve accuracy, but increases computational load and may lose robustness when a sensor’s observations are poor.Source: §4 · p. 7

4.1 Loosely-coupled Approaches

  • Loosely coupled methods estimate each sensor’s state independently before combining them, enabling modular adaptation and weighting when one sensor performs sub-optimally.Surveyed examples include EKF and ESKF fusion, INS-based initialization, and pose-graph optimization; the authors say these approaches improve on LiDAR-only accuracy but do not fully exploit sensor synergy.Source: §4 · p. 7, §4.1 · p. 7

4.2 Tightly-coupled Approaches

  • Tightly coupled estimation processes sensor measurements together, exploiting interdependencies between modalities in a unified state-estimation framework.IMU preintegration condenses measurements between keyframes into a factor, making it easier to include high-frequency IMU data in factor graphs.Source: §4 · p. 7, §4.2 · p. 8
  • Graph-based methods combine IMU preintegration with LiDAR factors, features, or global optimization to estimate odometry and trajectories.Examples include LIPS’s continuous IMU and plane factors, LIO-SAM’s use of preintegration for deskewing and graph factors, and GPU-accelerated matching with fixed-lag smoothing.Source: §4.2 · pp. 8–9
  • Filter-based methods target lower computational loads; FAST-LIO introduces a Kalman-gain formula, while FAST-LIO2 registers raw measurements to maps and uses an ikd-Tree.LINS still faced considerable computational load and slow processing; Faster-LIO replaces ikd-Tree with incremental voxels, and Inv-LIO2 is described as more accurate but slower than Inv-LIO1.Source: §4.2 · p. 9
  • Other tightly coupled methods target dynamic or degraded settings through dynamic-point handling, motion constraints, feature-quality measures, or adaptive keyframes.The surveyed examples include factor-graph approaches to dynamic scenes, ground-vehicle constraints, intensity and photometric features, and adaptive keyframe generation for extreme environments.Source: §4.2 · p. 10
  • High-frequency methods divide scans into subframes or process points individually, while learning-based and continuous-time methods address loop closure, asynchronous data, or nonlinear optimization.Point-LIO’s point-by-point framework is described as naturally eliminating motion distortion; CLINS uses cubic B-splines for asynchronous data, and DLIO uses a hierarchical geometrical observer.Source: §4.2 · pp. 9–10

5 Multiple LiDARs

Multiple LiDARs can address single-LiDAR coverage, interference, and scan-registration limitations, but synchronization, computational cost, and asynchronous measurements remain challenges.

5 Multiple LiDARs

  • Multiple LiDARs broaden scanning coverage, reduce interference, and can improve scan registration compared with relying on a single LiDAR with a non-repetitive scanning pattern.These benefits motivate exploring multi-LiDAR odometry to address limitations of single-LiDAR systems.Source: §5 · p. 11
  • M-LOAM aggregates features from synchronized LiDARs, while a decentralized EKF can handle asynchronous LiDARs but does not fully exploit their combined measurements.PPS synchronization adds hardware complexity; PTP may require additional effort to synchronize data.Source: §5 · p. 11
  • Independent state estimation from each LiDAR can let one sensor’s occlusion cascade into later estimates; discarding delayed scans improves robustness at the cost of information.M-LIO associates asynchronous measurements but does not compensate for their temporal discrepancies.Source: §5 · p. 11
  • IMU propagation compensates for temporal discrepancies among LiDARs, but interpolation can add errors and prolonged discrepancies increase reliance on noisy IMU estimates.MA-LIO uses B-spline interpolation and point-wise uncertainty; SLICT accumulates point clouds within a designated interval to keep accumulation duration consistent.Source: §5 · p. 11
  • Multiple-LiDAR systems still face placement optimization and higher computational demands, alongside issues inherited from single-LiDAR systems.Limited FOV, sensor interference, and irregular scanning patterns can hinder state estimation and scan registration.Source: §5 · p. 11

6 Fusion with Other Sensors

Sensor fusion can address environments where LiDAR measurements or geometry are challenging, using complementary modalities such as cameras, thermal cameras, radar, event cameras, and leg-contact information. The survey describes integration by pipeline stage and notes that fusion requires careful planning because it adds computational and sensor-specific challenges.

6 Fusion with Other Sensors

  • LiDAR odometry struggles in rain, snow, dust, and environments with sparse or repetitive geometry, motivating multimodal fusion.Scan-matching ambiguity in tunnels or highways can reduce state-estimation precision.Source: §6 · p. 12
  • RGB cameras add color and texture where geometric features are scarce; systems combine camera, LiDAR, and IMU measurements through filters, optimization, or jointly operating subsystems.Coco-LIC uses a non-uniform B-spline continuous-time trajectory and tightly couples LiDAR and camera data.Source: §6 · pp. 12–13
  • The pipeline classifies sensor integration as loosely or tightly coupled according to the stage where additional data is used; refined state estimation supports both odometry and mapping.The three stages are pre-processing, initial estimation, and state estimation.Source: Figure 3 · p. 12
  • Thermal cameras, radar, and event cameras offer alternatives to RGB cameras in challenging conditions, while VILENS also incorporates leg-contact information for legged-robot odometry.Rho et al. use stereo thermal cameras with LiDAR in indoor disaster scenarios.Source: §6 · p. 13
  • Sensor fusion can address individual-sensor limitations, but it adds computational demands and sensor-specific complexities, so indiscriminate fusion may not be optimal.The survey stresses planning and understanding each sensor’s requirements before deployment.Source: §6 · p. 13

7 Remaining Challenges

LiDAR odometry has advanced, but unresolved constraints remain across sensor data volume, motion compensation, heterogeneous devices, difficult environments, and multimodal integration. The survey identifies these issues as directions for further research.

7.1 LiDAR Inherent Problems

  • An OS1-128 can produce several 100K points per frame at up to 20 Hz, making real-time processing computationally demanding; multi-LiDAR optimization can take over 100 ms per sweep in two NTU VIRAL studies.That processing time may suit keyframe systems but is impractical when estimating every scan.Source: §7.1 · pp. 13–14
  • Motion during a scan causes distortion requiring de-skewing; continuous-time B-spline compensation can improve point transformations but raises computation, while LiDAR also faces limited FOV, sparse data, and installation trade-offs.A constant-velocity model may lack accuracy under aggressive motion or uncertain velocity estimates.Source: §7.1 · p. 14

7.2 Heterogeneous LiDARs

  • Mechanical and scanning solid-state LiDARs differ in viewing angles and scan patterns, and device-specific FOV and resolution can require algorithm adjustments; KISS-ICP is versatile, but generalized LiDAR-inertial and multimodal methods remain lacking.The survey identifies potential for improvement in generalized approaches.Source: §7.2 · p. 14

7.3 Degenerative Environment

  • Sparse or repetitive geometry can make scan matching ambiguous and pose estimates inaccurate; proposed responses include degeneracy-aware methods, sensor fusion, and additional intensity or reflectivity cues.These approaches include adaptive parameters, residual uncertainty handling, and loosely coupled sensor fusion.Source: §7.3 · pp. 14–15
  • FMCW LiDAR Doppler velocity objectives improve scan matching in feature-scarce environments, with DICP reporting high-accuracy odometry on a 900-meter tunnel sequence.Later work integrates Doppler velocity factors into continuous-time odometry.Source: §7.3 · p. 15

7.4 Degraded Environment

  • Weather can degrade LiDAR detection, and although denoising methods target weather noise, their performance in LiDAR odometry has not been sufficiently investigated; glass reflections also challenge scan matching, with no complete algorithmic remedy identified.The survey distinguishes degraded sensing conditions from degenerative environments.Source: §7.4 · pp. 15–16

7.5 Multi-Modal Sensors

  • Multimodal odometry faces difficult calibration, sensor-placement trade-offs, and asynchronous data frequencies; dedicated studies of placement and broader cross-modal synchronization approaches remain limited.Excessive LiDAR overlap can add redundant data and computational cost, potentially offsetting accuracy gains.Source: §7.5 · p. 16

8 Datasets & Evaluation

This section reviews public LiDAR datasets and ground-truth and evaluation methods, then benchmarks LiDAR-only against LiDAR-inertial odometry. Results indicate general robustness advantages for LiDAR-inertial methods, with important limitations on long sequences.

8.1 Public Datasets

  • Dataset coverage includes UAV, handheld, ship, and heterogeneous-LiDAR settings, while uniform hardware, controlled weather, and stationary vehicles leave important generalization gaps.LIBRE reports distance error for 12 LiDARs but uses one LiDAR per sequence and stationary vehicles in artificially controlled weather.Source: §8.1 · pp. 17–18
  • No single public dataset is universally comprehensive, so dataset selection must match the characteristics needed for robust, adaptable odometry development.Datasets vary in environment, location, weather, sensor configuration, and platform motion.Source: §8.1 · pp. 17–18

8.2.1 Groundtruth generation

  • Ground truth may come from RTK-compensated GPS, GPS/IMU/INS, SLAM, tracking systems, GCPs, or TLS, with suitability depending on signal availability and setup.Tracking systems are useful in controlled settings; SLAM can provide a reference where GPS is unavailable.Source: §8.2.1 · pp. 18–19

8.2.2 Evaluation Methods

  • ATE measures average deviation across corresponding trajectory poses, whereas RTE measures relative-pose deviation over shorter trajectory segments.ATE and RTE are typically calculated using RPG and EVO.Source: §8.2.2 · p. 19
  • Start-to-End Error compares odometry and registration-derived endpoint displacement; GCP alignment and entropy provide additional positioning and scan-registration assessments.Entropy tends to rise when poorly registered point clouds are combined; endpoint error is useful when reliable trajectory ground truth is difficult to obtain.Source: §8.2.2 · pp. 19–20, Figure 4 · p. 19

8.3 Benchmark Results

  • LiDAR-inertial odometry is generally more robust and helps on short dynamic paths, but is not invariably superior on long HeLiPR sequences prone to cumulative drift.The trajectory figure zooms in on LiDAR-only deviations from ground truth relative to LiDAR-inertial odometry in ConSLAM.Source: Table 4 · p. 20, Figure 5 · p. 20, §8.3 · p. 21

9 Conclusion

The survey reviews recent LiDAR odometry advances, their strengths and weaknesses, and challenges, future directions, datasets, and evaluation metrics. It concludes that approaches must be chosen for application needs because performance and resource demands trade off.

9 Conclusion

  • The survey covers recent LiDAR odometry advances, their strengths and weaknesses, remaining challenges, future directions, datasets, and evaluation metrics.Source: §9 · p. 21
  • LiDAR odometry has no one-size-fits-all solution; sensor configuration should match application accuracy, environment, and computational resources.The survey recommends LiDAR-only methods for well-defined settings with limited compute, and tightly coupled multisensor approaches when high accuracy across varied environments is required.Source: §9 · p. 21
  • The authors expect deep learning, multimodal sensor fusion, and continued sensor and algorithm development to contribute to more accurate and robust odometry.Source: §9 · p. 22
Loading 2312.17487v1…