Outlier-Resistant Extrinsic Calibration of High-Resolution LiDAR and Camera in Targetless Environments

In this article, we develop a novel method for calibrating the extrinsic parameters of a high-resolution LiDAR and the RGB camera in unstructured environments. The method eliminates the need for manual calibration objects and achieves calibration by aligning natural line features in the two sensors. The robust cost function called truncated least squares (TLS) is adopted to redefine the problem of estimating 3-D pose from 2-D–3-D line correspondences, which is known as the perspective-n-lines (PnL) problem. Therefore, the developed method can handle large-scale outliers and cope with cluttered features in the scene. Besides, the PnL-TLS formulations are given as a unified quadratically constrained quadratic problem (QCQP) using quaternions, which allows for elegant utilization of semidefinite solvers. In addition, the performance of the developed method is validated through simulations and real-world experiments. The simulation results demonstrate that the method performs on par with state-of-the-art approaches in the absence of outliers. The rotation error as well as the translation error can be reduced within 2° under the circumstances of 60% outliers in the original data. For real-world experiments, we conduct various analyses from different perspectives, including projecting point clouds onto images and coloring 3-D point clouds using image pixel values. The experimental results verify the effectiveness of the developed method in practical scenarios.