A Robust Online Miscalibration Detection and Correction Method for LiDAR-Camera

Abstract

The increasing reliance on multi-sensor systems to enhance the robustness and accuracy of robotic systems introduces new challenges, particularly in calibrating multiple sensors. Traditional offline calibration techniques, which use targets like checkerboards, assume that extrinsic parameters between sensors remain constant during operation. However, this assumption does not account for drift caused by external factors such as vibration and deformation over time. In this letter, we introduce CalibOnline, a novel method for the online detection and correction of miscalibration in multi-sensor setups. First, we propose a unified representation of LiDAR and camera data as depth maps to reduce calibration uncertainties due to data modality discrepancies. We then explore the robust characteristic of depth discontinuity edges to facilitate efficient matching between depth maps. The impact of extrinsic variations on edge-matching constraints is analyzed, leading to the design of a miscalibration detection module to monitor extrinsic parameters. Finally, we frame the correction of extrinsic parameters as an on-manifold optimization problem, enhancing the convergence of the estimated parameters. Experimental results across various scenarios demonstrate the robustness and accuracy of our proposed method. Code is open-sourced at: https://github.com/cchester25/CalibOnline.git.