Eigen Is All You Need: Efficient Lidar-Inertial Continuous-Time Odometry With Internal Association

Thien-Minh Nguyen, Xinhang Xu, Tongxing Jin, Yizhuo Yang, Jianping Li, Shenghai Yuan, Lihua Xie

Eigen Is All You Need: Turbocharging Continuous-Time Lidar Odometry

In the world of robotics, Lidar-Inertial Odometry (LIO) is the bread and butter of navigation. While traditional "discrete-time" systems estimate where a robot is at specific moments, the field is shifting toward "continuous-time" (CT) models. These models represent a robot's path as a smooth mathematical function (like a B-spline), which is far better at handling high-speed maneuvers and sensors that don't fire at the exact same time.

The catch? Continuous-time optimization is notoriously slow. In their new paper, researchers from Nanyang Technological University introduce SLICT2, a system that proves you don't need heavy-duty optimization libraries to achieve state-of-the-art results. By stripping back the complexity and focusing on "internal association," they’ve made CT-LIO faster and more accurate than ever.

TL;DR

SLICT2 is a continuous-time lidar-inertial odometry system that replaces heavy nonlinear solvers with a streamlined, iterative linear approach using the Eigen library. By re-calculating feature associations within the optimization loop, it achieves up to an 8x speedup over standard methods while maintaining superior accuracy in high-speed scenarios.

What Problem Does This Paper Address?

Continuous-time LIO (CT-LIO) is theoretically superior for high-dynamic motion because it treats the trajectory as a continuous curve rather than a series of "snapshots." However, the computational cost is a massive bottleneck.

Most existing systems rely on general-purpose Nonlinear Least Squares (NLS) solvers like Ceres or GTSAM. While powerful, these solvers carry significant "bookkeeping" overhead—building the problem, managing Jacobians, and ensuring every step decreases the cost function. In a real-time robotics application, this overhead often prevents the use of dense control points, forcing researchers to choose between speed and trajectory smoothness.

Methods

The authors propose a "straightforward" approach that challenges conventional wisdom. Their key innovations include:

  • The "Solve-Associate-Solve" Loop: Standard methods usually find matches (associations) between lidar points and the map once per iteration. SLICT2 performs Internal Association, re-doing the feature-to-map association immediately after each incremental optimization step.
  • Linear Solver over NLS: Instead of using a complex NLS solver, they implement a customized on-manifold solver based on the Eigen library. They found that the CT-LIO problem is "well-behaved" enough to converge in just 2–3 iterations if the associations are updated frequently.
  • B-Spline Trajectory: They use a B-spline formulation (order 4) to represent the trajectory, which naturally handles gravity, IMU biases, and high-frequency motion without the need for "pre-integration" tricks used in discrete systems.
  • Parallelization: The system is built from the ground up to calculate residuals and Jacobians for thousands of lidar points in parallel, bypassing the architectural bottlenecks of general-purpose solvers.

Datasets

The authors evaluated SLICT2 against several state-of-the-art (SOTA) methods (FAST-LIO2, CT-ICP, CLIC, and the original SLICT) using:

  • NTU VIRAL: A dataset from an aerial vehicle equipped with a 16-channel lidar and high-frequency IMU, featuring centimeter-level ground truth.
  • Multi-Campus Dataset (MCD): A challenging dataset featuring high-velocity sequences and non-repetitive (Livox) lidar sensors.

Key Insights and Ideas

  • Association > Iterations: The "aha moment" of the paper is that the correctness of the feature association matters much more than the number of optimization iterations. By updating associations inside the loop, the solver reaches a better solution much faster.
  • NLS Overhead is the Enemy: General-purpose solvers spend a surprising amount of time on internal management. By using a "raw" linear solver (Eigen) and a custom parallelized structure, the authors eliminated the overhead that makes other CT-LIO systems sluggish.
  • Convergence Speed: While a solver like Ceres might take 7–8 iterations to converge, SLICT2 typically converges in just 2 or 3 because the first iteration usually moves the state estimate close enough to significantly improve the next round of feature matching.

Key Results

  • Efficiency: SLICT2 is significantly faster than NLS-based approaches. In head-to-head timing tests, the BSU (Build-Solve-Update) process was roughly 8 times faster than a comparable implementation using the Ceres solver.
  • Accuracy: On the NTU VIRAL dataset, SLICT2 consistently produced the lowest or second-lowest Absolute Position Error (APE).
  • High-Speed Performance: In the MCD dataset (high-velocity UAV motion), SLICT2 outperformed FAST-LIO2 and CT-ICP, proving that its dense B-spline representation captures fast dynamics that discrete-time or sparse-control-point systems miss.
  • Real-Time Capability: The system maintains a mean processing time of ~83ms per loop, comfortably fitting within real-time requirements even with a high density of control points (0.01s knot length).

Limitations and Future Work

While SLICT2 is a major step forward for efficiency, the authors note a few areas for growth:

  • Feature Scope: The current implementation focuses on the core LIO pipeline. The authors suggest that future versions could incorporate online temporal offset estimation (time-sync between sensors) and extrinsic calibration.
  • Multi-Modal Fusion: The framework is currently lidar-inertial; extending this "internal association" logic to visual features or other sensors remains an open challenge.
  • Hardware Constraints: While highly efficient on a PC (i9 CPU), testing on lower-power embedded platforms (like those found on small drones) would be the ultimate test of its "efficiency-first" philosophy.

Summary updated

Abstract

In this letter, we propose a continuous-time lidar-inertial odometry (CT-LIO) system named SLICT2, which promotes two main insights. One, contrary to conventional wisdom, CT-LIO algorithm can be optimized by linear solvers in only a few iterations, which is more efficient than commonly used nonlinear solvers. Two, CT-LIO benefits more from the correct association than the number of iterations. Based on these ideas, we implement our method with a customized solver where the feature association process is performed immediately after each incremental step, and the solution can converge within a few iterations. Our implementation can achieve real-time performance with a high density of control points while yielding competitive performance in highly dynamical motion scenarios. We demonstrate the advantages of our method by comparing with other existing state-of-the-art CT-LIO methods.
BibTeX
@article{nguyen2024eigen,
  author = {Thien-Minh Nguyen and Xinhang Xu and Tongxing Jin and Yizhuo Yang and Jianping Li and Shenghai Yuan and Lihua Xie},
  title = {{Eigen Is All You Need: Efficient Lidar-Inertial Continuous-Time Odometry With Internal Association}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2024},
  url = {https://openreview.net/forum?id=ilOGfU0ldE},
}

Loading related content...

Figure