Self-Calibration of Multiple LiDARs for Autonomous Vehicles

Abstract

We propose a novel LiDAR extrinsic calibration method that does not require external calibration settings or human intervention. Instead of using fixed calibration targets, our algorithm automatically detects valid geometry-level features in random surrounding scenes. As an assumption-free approach, we introduce an initial state estimation algorithm to aid the global optimization process. This is achieved by extending the Rotation-Invariant Feature (RIF) to tolerate noise and uncertainty from LiDAR and environments, which results in robust point correspondence matching. In order to improve efficiency, the translation and rotation offsets are optimized separately through the Branch-and-Bound (BnB) optimization. The proposed algorithm shows promising results on an outdoor calibration dataset and outperforms previous calibration methods. It also demonstrates comparable robustness to offline calibration approaches, but with fewer constraints and assumptions.