Scene-Aware Online Calibration of LiDAR and Cameras for Driving Systems

Abstract

This paper introduces a robust method for accurately perceiving failures and calibrating multi-line LiDARs and cameras in natural environments in an online setting. Traditional target-free calibration methods rely on matching the spatial structures of 3D point clouds with image features. However, obtaining dense point cloud data in a short amount of time for matching and optimization is challenging in online applications. To address this, our method uses single-frame sparse LiDAR point clouds for robust feature extraction and matching, with further optimization through contextual observation. Moreover, our approach is capable of perceiving and re-calibrating extrinsic errors in online natural scenes, thus enhancing the calibration’s robustness. We demonstrate the robustness and generalizability of our method using our own datasets LIVOX-Road, with evaluation results indicating subpixel accuracy. The code is released at https://github.com/JMU-Robot/LiDAR-Camera-Online-Calibration.