A proposed calibration method for the laser scanner and inertial sensor on a ground vehicle handled inclined surfaces better than a comparison method in tests. On Husky ground-vehicle sequences collected on sloping ground, GRIL-Calib produced an incorrect calibration, while the proposed method converged reasonably close to measured values and to the flat-ground mean extrinsics.
The approach is designed for planar vehicle motion on non-flat ground. It uses continuous-time optimization, which models the vehicle’s motion as changing continuously, together with two extra checks: one based on distance to the ground plane and another based on the ground’s inclination.
A mixed picture across datasets
The researchers evaluated the method on a Husky ground-vehicle dataset, the M2DGR dataset and an offroad vehicle dataset. Each dataset used between three and seven calibration sequences.
On flat-ground Husky data, the proposed method was more repeatable across all extrinsic parameters, with the clearest gains in yaw, the vehicle’s rotation around the vertical axis, and in its x-y position estimates.
The M2DGR results were less uniform. Repeatability improved for x-y position but was slightly worse for yaw. Both methods showed a yaw standard deviation of about one degree, meaning their estimates spread by roughly that amount across sequences.
The offroad results again favored the proposed method on extrinsic spread, or how much the estimated sensor relationship varied from run to run, except for a slight worsening in the y-component of orientation error.
The calibration still depends on how it starts
Additional tests showed that the choice of residuals, the error terms the optimizer tries to reduce, mattered. In an offroad ablation, versions using flat-ground residuals had significantly greater spread in position estimates, while OA-Calib had the smallest spread among the proposed methods.
The experiments also exposed sensitivity to initialization, meaning the starting sensor estimates supplied to the optimizer. When those estimates were deliberately misinitialized, cases using the distance residual showed much larger spread in position-related parameters. Without that residual, the vertical sensor-offset estimate, or z-extrinsic, was not refined from its starting value and had an error of around 25 centimeters.
The study assessed calibration quality by the spread of estimates across sequences, treating lower standard deviation as better repeatability. It also compared the mean calibration with measured extrinsics as a sanity check.
What the tests do not establish
The paper compares the proposed algorithm with GRIL-Calib; GCT-Calib was not included because its source code was unavailable. The experiments assess calibration repeatability and agreement with measured reference values, not whether a vehicle later localizes, maps or navigates better.
The evaluation remains limited to the tested planar-motion settings, with three to seven sequences per dataset. The method also requires the IMU height to be known, uses a gravity direction resolved on flat ground to calculate inclination, and assumes that the extracted ground surface is reasonably planar.
The work is an arXiv preprint dated 25 August 2026. The paper states that its implementation and experiments are open-sourced.
Paper data and sources
Original title: Extending Ground-Constraint LiDAR-IMU Calibration to Tilted Surfaces in a Continuous-Time Framework
Authors: Vassili Korotkine, Pierre Chamoun, Mohammed Ayman Shalaby, James Richard Forbes
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text