Abstract
In this paper, 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. The source code will be released for the benefit of the community.
Introduction
Lidar-Inertial Odometry (LIO) is a cornerstone of robotic navigation. While traditional discrete-time systems estimate states at specific timestamps, continuous-time (CT) models represent trajectories as smooth mathematical functions, such as B-splines. This approach is inherently better at handling high-speed maneuvers and asynchronous sensor data. However, CT-LIO optimization is computationally expensive, often requiring heavy-duty nonlinear least squares (NLS) solvers like Ceres or GTSAM. These solvers introduce significant bookkeeping overhead, which limits real-time performance and the density of control points. SLICT2 addresses these bottlenecks by replacing general-purpose NLS solvers with a streamlined, iterative linear approach using the Eigen library, centered on the principle of "internal association."
Methodology
Continuous-Time Trajectory Representation
SLICT2 utilizes a 4th-order B-spline formulation to represent the robot's trajectory. Unlike discrete-time systems that rely on IMU pre-integration between states, the continuous-time model naturally incorporates gravity and IMU biases. This allows for a more accurate representation of high-frequency motion by treating the trajectory as a continuous curve rather than a series of snapshots.
The "Solve-Associate-Solve" Loop
The core innovation of SLICT2 is the Internal Association mechanism. Standard CT-LIO methods typically perform feature association (matching lidar points to the map) once per optimization iteration. SLICT2, however, re-calculates these associations immediately after each incremental optimization step within a customized on-manifold solver.
By updating associations inside the loop, the system ensures that the solver always works with the most accurate correspondences. The authors found that the CT-LIO problem is sufficiently well-behaved to converge in just 2–3 iterations when associations are updated frequently, eliminating the need for the complex step-size management and cost-function monitoring found in general NLS libraries.
Experimental Results
Efficiency and Real-Time Performance
SLICT2 was evaluated against state-of-the-art methods including FAST-LIO2, CT-ICP, and CLIC. In timing tests, the Build-Solve-Update (BSU) process in SLICT2 was approximately 8 times faster than a comparable implementation using the Ceres solver. The system maintains a mean processing time of ~83ms per loop, enabling real-time operation even with a high density of control points (0.01s knot length).
Accuracy in Dynamic Scenarios
The system was tested on the NTU VIRAL and Multi-Campus Dataset (MCD). On NTU VIRAL, SLICT2 consistently achieved the lowest or second-lowest Absolute Position Error (APE). In the high-velocity UAV sequences of the MCD dataset, SLICT2 outperformed discrete-time and sparse-control-point systems, demonstrating that its dense B-spline representation effectively captures fast dynamics that other models miss.
Conclusion
SLICT2 demonstrates that for continuous-time lidar-inertial odometry, the correctness of feature association is more critical than the complexity of the optimization solver. By leveraging a raw linear solver (Eigen) and a custom parallelized structure for internal association, SLICT2 removes the overhead of general-purpose NLS libraries. This results in a system that is both significantly faster and more accurate in highly dynamic environments. Future work will explore extending this internal association logic to multi-modal sensor fusion and online extrinsic calibration.
Summary updated