Source-linked AI summary
LINS: A Lidar-Inertial State Estimator for Robust and Efficient Navigation
Chao Qin, Haoyang Ye, Christian E. Pranata, Jun Han, Shuyang Zhang, Ming Liu
TL;DR
LINS addresses the computational cost and environmental sensitivity of lidar-based ego-motion estimation by tightly fusing lidar with an IMU. It uses an iterated ESKF and robocentric state formulation, achieving comparable accuracy and stability to state-of-the-art lidar-inertial odometry with substantially lower runtime.
Problem
Lidar navigation is sensitive to surroundings, motion distortion, and sparse point clouds, while graph-optimization lidar-inertial odometry is too computationally expensive for real-time navigation.
Method
LINS tightly couples a 3D lidar and 6-axis IMU using an iterated ESKF with a robocentric local-frame formulation for efficient, stable ego-motion estimation.
Results
LINS outperforms lidar-only methods and achieves comparable accuracy and stability to state-of-the-art lidar-inertial odometry at much lower computational cost.
Takeaways & Limitations
LINS provides real-time, long-term, robust, and high-precision ego-motion estimation across challenging environments.
Abstract
from arXiv · showhide
We present LINS, a lightweight lidar-inertial state estimator, for real-time ego-motion estimation. The proposed method enables robust and efficient navigation for ground vehicles in challenging environments, such as feature-less scenes, via fusing a 6-axis IMU and a 3D lidar in a tightly-coupled scheme. An iterated error-state Kalman filter (ESKF) is designed to correct the estimated state recursively by generating new feature correspondences in each iteration, and to keep the system computationally tractable. Moreover, we use a robocentric formulation that represents the state in a moving local frame in order to prevent filter divergence in a long run. To validate robustness and generalizability, extensive experiments are performed in various scenarios. Experimental results indicate that LINS offers comparable performance with the state-of-the-art lidar-inertial odometry in terms of stability and accuracy and has order-of-magnitude improvement in speed.
I. INTRODUCTION
LINS addresses the need for robust, real-time ego-motion estimation by tightly fusing lidar and IMU measurements with an iterated ESKF and robocentric state formulation. The proposed system targets challenging environments while improving computational efficiency over graph-optimization-based lidar-inertial odometry.
- Motivation: Lidar-based navigation benefits from wide horizontal FOV and lighting invariance but remains sensitive to surroundings, motion distortion, and sparse point clouds.These limitations become especially problematic in challenging settings such as wide, open areas.
- Motivation: IMU fusion supplies high-frequency short-term motion constraints and helps recover point clouds affected by dynamic motion distortion.However, graph-optimization-based lidar-inertial odometry can require more than 100 milliseconds per scan, limiting real-time use.
- Proposed method: LINS is a lightweight lidar-inertial state estimator designed for real-time navigation of unmanned ground vehicles.Its core design uses tightly coupled lidar-inertial odometry and an iterated error-state Kalman filter.
- Proposed method: A robocentric iterated ESKF shifts the local reference frame at every lidar time-step and uses relative pose estimates to update the global pose.The formulation is intended to support long-term stability and is evaluated across challenging scenarios.
- Contributions: LINS is presented as a tightly coupled 6 DOF lidar-inertial estimator that uses iterated Kalman filtering and provides online source code.The introduction identifies it as the first tightly coupled LIO claimed to solve 6 DOF ego-motion through iterated Kalman filtering.
- System scope: The system pipeline includes feature extraction, lidar-inertial odometry, and mapping, with the paper focusing on the lidar-inertial odometry module.The remaining paper is organized around related work, system overview, experiments, and conclusion.
A. Lidar-only Methods
Prior lidar-only and fusion methods estimate motion through scan matching or separate sensor processing, while tightly coupled approaches jointly optimize lidar and inertial constraints. LINS focuses on iterated Kalman filtering to address nonlinear matching and linearization problems in this setting.
- Lidar-only methods: Lidar-only approaches commonly use ICP-based scan matching, while LOAM and LeGO register edge and planar features against incrementally built maps.LeGO adapts LOAM for unmanned ground vehicles and filters unreliable features using ground extraction and segmentation.
- Fusion methods: Loosely coupled fusion processes lidar and IMU constraints separately before combining them, which is computationally efficient but can lose information.IMU-aided LOAM and related methods use inertial estimates as priors or combine them with lidar-based pose estimates.
- Fusion methods: Tightly coupled approaches directly fuse lidar and inertial measurements through joint optimization or extended Kalman filtering.Existing methods include optimization-based systems such as LIPS and LIOM, as well as EKF-based approaches using lidar constraints.
- Limitations of prior work: Graph optimization in LIOM uses inertial pre-integration and lidar plane constraints, but local-map batch optimization is too time-consuming for real-time applications.A 2D lidar-aided EKF also has limited scope because it requires surrounding planes to have an orthogonal structure.
- LINS approach: EKF linearization errors can cause poor performance or divergence, especially when incorrect initial poses produce wrong lidar feature matchings.LINS addresses this with iterated Kalman filtering that repeatedly seeks better matchings and an error-state representation.
- System overview: The system overview defines feature extraction, lidar-inertial odometry, and mapping as its three major modules for estimating 6 DOF motion and building a global map.The LIO module propagates and updates the state using IMU measurements and features from consecutive scans.
- Scope: The paper focuses on the odometry module and refers readers elsewhere for detailed feature-extraction and mapping procedures.The feature-extraction implementation is also delegated to prior work.
C. Lidar-Inertial Odometry with Iterated ESKF
The LIO module uses a robocentric iterated ESKF to estimate relative motion from consecutive lidar features and IMU measurements while limiting linearization errors from growing uncertainty.
- C. Lidar-Inertial Odometry with Iterated ESKF: The LIO module estimates relative vehicle transformation from IMU measurements and features extracted from two consecutive lidar scans.The system uses the previous lidar-time IMU-affixed frame as its local frame.
- C. Lidar-Inertial Odometry with Iterated ESKF: The robocentric formulation represents the state relative to a moving local frame to prevent large linearization errors caused by growing uncertainty.The local frame is the IMU-affixed frame at the previous lidar timestep.
- C. Lidar-Inertial Odometry with Iterated ESKF: The local state includes relative translation and rotation, velocity, acceleration bias, gyroscope bias, and local gravity.The translation and rotation describe the transformation between consecutive local IMU frames.
- C. Lidar-Inertial Odometry with Iterated ESKF: An error-state representation uses a correction vector with a 3 DOF angle error, after which the corrected state is recovered through the boxplus operation.Quaternion multiplication and exp: R3 → SO(3) represent rotational composition.
2) Propagation:
The propagation step advances the robocentric state, error state, and covariance whenever new IMU measurements arrive, using a linearized continuous-time model and its discretization.
- 2) Propagation:: When a new IMU measurement arrives, the system propagates the error state, covariance matrix, and state prior.The propagated quantities are the error state δx, covariance Pk, and state prior.
- 2) Propagation:: The linearized IMU error-state model uses an error-state transition matrix, a noise Jacobian, and a Gaussian noise vector.Ft is the transition matrix and Gt is the noise Jacobian at time t.
- 2) Propagation:: The model removes bias and gravity effects from raw accelerometer and gyroscope measurements to obtain acceleration and angular-rate inputs.The corrected quantities are calculated from raw sensor measurements.
- 2) Propagation:: Discretizing the continuous-time model produces propagation equations over consecutive IMU timesteps using Δt and calibrated noise covariance Q.Q is computed offline during sensor calibration.
- 2) Propagation:: The discrete-time robocentric propagation model integrates IMU measurements to advance the local state.The paper refers to prior work for details of robocentric-state IMU integration.
3) Update:
The update step uses an iterated Kalman scheme whose residuals come from point-edge and point-plane correspondences, with new matches recomputed at every iteration.
- 3) Update:: The iterated update scheme is identified as the primary contribution of the work.It links iterated Kalman filtering with an optimization problem based on deviation from the prior state.
- 3) Update:: The measurement model stacks residuals from point-edge and point-plane pairs and evaluates them with a Mahalanobis norm.Jk is the measurement-noise Jacobian and Mk is the measurement-noise covariance.
- 3) Update:: For edge features, the residual measures distance to a corresponding edge; for planar features, it measures distance to a plane formed by three points.The correspondences are selected according to the feature type.
- 3) Update:: Each iteration finds new matched edges and planes, then recomputes the Jacobians and Kalman gain to further minimize the error metric.The correction vector is indexed by iteration, and the update terminates when the residual falls below a threshold.
- 3) Update:: After the state correction, raw distorted features are undistorted using the estimated relative transformation before the next state is initialized.The covariance retains uncertainty for velocity, biases, and local gravity, while relative-pose covariance is set to zero.
4) State Composition:
After each update, the robocentric formulation composes the local motion with the global pose to update the global state.
- 4) State Composition:: The global pose is updated through a composition step whenever the robocentric update finishes.This composition maps the updated local estimate into the global frame.
5) Initialization:
LINS initializes its filter using offline calibration, stationary IMU measurements, and unbiased acceleration before motion.
- Initial acceleration bias and lidar-IMU extrinsic parameters come from offline calibration, while gyroscope bias uses stationary measurements.
- Initial roll and pitch are obtained from unbiased acceleration measurements before the vehicle moves.
- The robocentric formulation facilitates initialization of the filter state.
IV. EXPERIMENTS
The experiments compare LINS with LeGO, LOAM, and LIOM across scenarios, evaluating both map-refined and pure odometry on identical computing infrastructure.
- LINS is evaluated against LeGO, LOAM, and LIOM on a 2.4GHz quad-core laptop with 8GiB memory.
- The experiments use C++ implementations running under ROS on Ubuntu Linux.
- LINS uses the LeGO mapping algorithm for its mapping module.
- Evaluation distinguishes map-refined odometry (MRO) from pure odometry (PO) to account for both mapping-refined and initial trajectories.
A. Indoor Experiment
In an indoor parking-lot test without ground truth, LINS produced a pure-odometry trajectory that aligned closely with its map-refined trajectory, unlike competing methods.
- The indoor experiment used a bus-mounted RS-LiDAR-16 and an IMU in a parking lot.
- LINS-PO aligned precisely with LINS-MRO in the indoor test.The passage characterizes MRO as generally almost drift-free indoors and more accurate than PO.
- LeGO-PO and LIOM-PO exhibited noticeable yaw-angle drift.
B. Large-scale Outdoor Environment
LINS was evaluated in varied outdoor environments and showed low drift, including competitive performance with LIOM and successful operation in feature-sparse urban scenes.
- Outdoor scenarios: LINS was tested in city, port, industrial-park, and forest environments, with maps generated for representative scenes.
- Evaluation setup: The outdoor evaluation measured position gaps against GPS-generated ground truth.
- Drift performance: LINS-MRO achieved 1.56% relative drift, compared with 1.40% for LIOM; LINS-PO achieved 2.75%.
- Trajectory and mapping results: LINS produced trajectories close to ground truth and maps with higher fidelity than LeGO.
2) Urban Experiment:
In the urban experiment, LINS remained stable in a feature-poor scene and achieved accuracy close to LIOM while substantially reducing runtime. Its performance was comparable to state-of-the-art lidar-inertial odometry at much lower computational cost.
- Urban Experiment: 56 edge features per scan was the lowest among all tested scenarios, yet LINS stably followed the real-world road in the feature-less urban scene.The experiment used GPS positions as ground truth.
- Urban Experiment: 1.79% MRO drift and 4.42% PO drift were obtained by LINS, close to LIOM’s 1.76% and 4.44%, respectively.ATE comparisons likewise showed LINS close to LIOM in accuracy.
- Runtime Comparison: LINS required less than 30 ms per scan, whereas LIOM required more than 100 ms and reached 223 ms in the feature-abundant parking lot.The runtime comparison demonstrates substantially better real-time capacity for LINS.
- Conclusion: LINS combines an iterated ESKF with a robocentric formulation to provide real-time, long-term, robust, and high-precision ego-motion estimation across challenging scenarios.The paper reports validation in city, port, industrial park, forest, and indoor parking-lot environments.
- Conclusion: LINS outperformed lidar-only methods and achieved comparable performance to state-of-the-art lidar-inertial odometry with much lower computational cost.This conclusion summarizes the reported stability, accuracy, and efficiency results.