GPIS-Based Calibration for Non-Overlapping Dual-LiDAR Systems Using a 2.5D Calibration Framework

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Extrinsic calibration of dual-LiDAR systems is crucial for autonomous driving.
• Non-overlapping fields of view (FoV) present significant challenges for traditional calibration methods.
• Correspondence-based techniques are often unreliable in scenarios with limited spatial overlap.

Purpose of the Study:

• To develop an engineering-oriented 2.5D calibration framework for dual-LiDAR systems.
• To address the challenges of extrinsic calibration in non-overlapping FoV configurations.
• To improve the accuracy, robustness, and feasibility of dual-LiDAR calibration.

Main Methods:

• A motion-guided planar alignment approach estimates initial horizontal extrinsics (x, y, yaw).
• Gaussian Process Implicit Surfaces (GPIS) are employed for refining extrinsics using spatially disjoint scans.
• The framework avoids calibration targets and reduces reliance on strong scene assumptions.

Main Results:

• Achieved centimeter-level lateral accuracy and sub-degree yaw error in high-fidelity simulations.
• Demonstrated consistent outperformance against motion-based and Bird's-Eye View (BEV)-based baselines under various noise conditions.
• Preliminary nuScenes study showed improved yaw accuracy and competitive lateral precision in a simulated non-overlapping dual-LiDAR setup.

Conclusions:

• The proposed 2.5D calibration framework offers a practical solution for non-overlapping dual-LiDAR systems.
• The method provides a favorable balance of accuracy, robustness, and engineering feasibility.
• It serves as an effective refinement stage for enhancing dual-LiDAR calibration.