A Tightly Coupled LiDAR-Inertial SLAM for Perceptually Degraded Scenes

Lin Yang, Hongwei Ma, Yan Wang, Jing Xia, Chuanwei Wang2022 · DOI

TL;DR

This paper introduces a novel tightly coupled LiDAR-Inertial SLAM system designed for challenging environments with limited visual features, such as tunnels and corridors. By combining an iterative Kalman filter for front-end odometry with pose graph optimization incorporating loop detection and ground constraints in the back end, the system achieves robust and accurate 6DOF state estimation and mapping.

What Problem Does This Paper Address?

Simultaneous Localization and Mapping (SLAM) is crucial for autonomous robot navigation, enabling robots to build maps of their surroundings while simultaneously tracking their own position within those maps. While LiDAR-based SLAM offers accurate range information and is independent of lighting conditions, it struggles in "perceptually degraded scenes" – environments lacking distinct geometric features, such as long, uniform corridors or underground tunnels. In these scenarios, LiDAR can suffer from motion distortion, low-frequency updates, and sparse point clouds, leading to inaccurate feature matching and significant pose drift. Fusing LiDAR with Inertial Measurement Units (IMUs) can compensate for these weaknesses, as IMUs provide high-frequency pose estimates. However, IMUs are prone to drift over time due to noise and bias. Existing loosely coupled fusion methods often treat sensor measurements as black boxes, leading to information loss. Tightly coupled methods, which fuse raw sensor data, offer better robustness and accuracy but can be computationally intensive. This paper aims to bridge this gap by proposing a tightly coupled LiDAR-IMU SLAM system specifically optimized for perceptually degraded scenes, balancing accuracy, real-time performance, and robustness.

Methods

The proposed SLAM system is built upon a tightly coupled LiDAR-IMU fusion framework, comprising two main components: a front-end iterative Kalman filter for LiDAR-Inertial Odometry (LIO) and a back-end pose graph optimization.

  • Front-End (Iterative Kalman Filter): This component focuses on robust state estimation. It utilizes an iterative Kalman filter that incorporates both LiDAR and IMU measurements. The core idea is to propagate the robot's prior position and attitude using predictions and observations, enhancing attitude accuracy and system robustness. An "error-state" representation is employed to ensure the validity of linearization, and an iterative approach is used to refine matches and minimize errors, addressing potential issues with linearization errors in standard Extended Kalman Filters (EKFs).
  • Back-End (Pose Graph Optimization): This component refines the trajectory and builds a globally consistent map. It employs a keyframe selection strategy to manage computational load in large-scale environments. Crucially, it incorporates:
  • Loop Detection: Identifying when the robot revisits a previously mapped area to correct accumulated drift.
  • Ground Constraints: Utilizing detected ground planes to provide additional constraints on the robot's vertical movement, particularly useful in environments where horizontal features are scarce. These constraints are integrated into a pose graph optimization framework, aiming to minimize the overall error and ensure map consistency.

Datasets

The authors evaluated their algorithm using:

  • Public Dataset 1: A six-set dataset designed to compare performance across varying speeds (fast-to-slow) for both translation and rotation. This dataset provides ground truth from a high-precision motion capture instrument.
  • Public Dataset 2 (KITTI 2011_09_30_drive_0028): A sequence from the well-known KITTI dataset, used to evaluate long-term performance and cumulative error. Ground truth is provided by outdoor high-precision GPS.
  • Self-Picked Datasets: Three datasets were recorded by the authors using their own experimental platform in three different perceptually degraded environments (corridor scenes). These were used for qualitative analysis.

Key Insights and Ideas

  • Iterative Kalman Filter for Robustness: The use of an iterative Kalman filter in the front-end is a key innovation. Unlike standard EKFs, this iterative approach can better handle linearization errors and improve the accuracy of state estimation, especially in challenging scenarios where feature extraction might be less reliable.
  • Tightly Coupled Fusion for Perceptually Degraded Scenes: The paper strongly advocates for tightly coupled LiDAR-IMU fusion as the superior approach for environments with sparse features. By directly fusing raw sensor data, it avoids the information loss inherent in loosely coupled methods.
  • Synergy of Front-End and Back-End: The system's strength lies in the synergy between its front-end and back-end. The robust odometry from the iterative Kalman filter provides a solid foundation for the pose graph optimization, which then corrects drift and ensures global consistency through loop detection and ground constraints.
  • Ground Constraints as a Critical Factor: The ablation studies highlight the significant impact of ground constraints. They are shown to effectively correct Z-axis drift, leading to a more globally consistent and accurate map, which is vital for robots operating on uneven terrain or in environments where vertical alignment is crucial.

Key Results

  • Accuracy: The proposed algorithm consistently achieved lower Root Mean Square Errors (RMSE) for both translational and rotational states compared to baseline methods (LOAM, LIO-Raw, LIO-No-Ext, LIO) across various speed datasets. For example, in the "Slow_1" dataset, the proposed method achieved an RMSE of 0.0248 m for translation, significantly better than others.
  • Robustness to Speed: The algorithm's performance was not significantly affected by changes in motion speed, outperforming the LIO baseline which showed performance degradation at higher speeds.
  • Long-Term Performance: Quantitative analysis on the KITTI dataset showed smaller translation and rotation errors for the proposed algorithm compared to Aloam, Lego_Loam, and Lio_Sam, indicating better long-term accuracy and reduced cumulative error.
  • Qualitative Mapping: In perceptually degraded scenes, the proposed method generated more complete and realistic geometric maps compared to other algorithms. While Aloam performed well, it showed some mismatching. Lego_Loam struggled with Z-axis drift, and Lio_Sam failed in highly similar scenarios due to incorrect loop detection.
  • Real-time Performance: The system demonstrated real-time capabilities, with average running times for key modules (front-end pre-processing, scan matching, map construction, loop detection, pose graph optimization) well within the limits for a 10 Hz operating frequency.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the algorithm is designed for perceptually degraded scenes, the exact true values for such environments are often unavailable, necessitating reliance on public datasets with ground truth and qualitative analysis for self-collected data.
  • The performance of loop detection relies on high correlation of correct data; incorrect matching can still lead to system failure, as observed with Lio_sam in similar scenarios.
  • Future Work:
    • The authors plan to extend their research to multi-sensor fusion, incorporating vision alongside LiDAR and IMU, and exploring other practical scenarios of perceptual degradation.
  • Further investigation into the structure of stored point cloud map data is proposed, aiming to improve map-based positioning and navigation for robots and ensure real-time performance.

Summary updated

BibTeX
@misc{yang2022tightly,
  author = {Lin Yang and Hongwei Ma and Yan Wang and Jing Xia and Chuanwei Wang},
  title = {{A Tightly Coupled LiDAR-Inertial SLAM for Perceptually Degraded Scenes}},
  year = {2022},
  doi = {10.3390/s22083063},
}
Figure