Dynamic Object-Based Point Cloud Registration for Roadside LiDAR Calibration in Low-Overlap Outdoor Deployments

Abstract

Accurate calibration of roadside light detection and ranging (LiDAR) sensors is essential for providing autonomous vehicles with beyond-line-of-sight perception capabilities in large-scale outdoor environments. However, in such scenarios, low intersensor overlaps and sparse point clouds result in unstable feature correspondences and insufficient geometric constraints, making accurate LiDAR calibration highly challenging. To address these limitations, we propose a low-overlap roadside LiDAR calibration framework based on dynamic vehicle point cloud registration. Specifically, we introduce a dynamic vehicle association method that matches vehicles observed by different LiDARs using the spatiotemporal consistency of their trajectories. Based on this, we propose a bi-level cross-spatiotemporal registration method. Cross-time trajectory registration provides an initial pose estimation, while cross-space vehicle-surface registration enhances surface completeness and point density, thereby improving calibration accuracy. Both field experiments and simulation studies demonstrate the effectiveness of the proposed method. In real-world roadside experiments, the proposed framework achieves a relative translation error (RTE) within 0.45 m and a relative rotation error (RRE) below 0.65° across different deployment scenarios. Simulation experiments further show that the method maintains stable performance under varying intersensor spacing and overlap conditions. These results highlight the potential of our approach to advance roadside LiDAR calibration, supporting large-scale continuous traffic perception in intelligent transportation systems.