Robust Extrinsic Calibration for LiDAR-Camera Systems via Depth and Height Complementary Supervision Network

Abstract

Extrinsic calibration between cameras and LiDAR (Light Detection and Ranging) is a fundamental step for numerous autonomous driving applications, such as 3D object detection, lane detection, and trajectory planning. However, conventional calibration methods are laborious and require dedicated data collection. Although recent research has shown the potential of learning-based solutions for end-to-end use, existing methods tend to focus on feature matching rather than geometric constraints, leading to inefficiency and instability in traffic scenarios with various disturbances. In this paper, we introduce an improved deep-learning-based joint calibration framework for LiDAR-camera systems, termed the Complementary Calibration Network (Co-CalibNet). The main contribution of our work lies in proposing a novel dual-channel geometric supervision calibration framework that integrates both depth and height supervision to achieve robust extrinsic calibration. Additionally, we introduce an Attention-based Fusion Module (AFM) for efficient feature fusion. Furthermore, we incorporate time compensation and iterative calibration techniques to further enhance the robustness of the algorithm in handling initial alignment errors. Evaluations on the DAIR-V2X and KITTI datasets demonstrate that Co-CalibNet achieves state-of-the-art calibration performance while exhibiting greater robustness to initial misalignment. Since it requires no targets or human effort, Co-CalibNet can be seamlessly integrated into any LiDAR-camera architecture, suggesting significant value and broad application prospects in autonomous driving systems.