Robust and High-Precision Point Cloud Registration Method Based on 3D-NDT Algorithm for Vehicle Localization

Point cloud registration plays a pivotal role in LiDAR-based vehicle localization, as its robustness and accuracy directly affect map quality and localization precision. The 3D Normal Distributions Transform (3D-NDT) is a competitive algorithm that performs well in noisy and dynamic environments. However, its effectiveness is limited by local feature blurring caused by voxelization. To address this issue, this paper proposes an improved 3D-NDT registration method that incorporates normal vector segmentation and refined voxelization techniques to improve registration accuracy and matching range. The process begins with plane point cloud clustering for initial voxel partitioning, followed by subdivision to ensure uniform voxel cell sizes, enabling appropriate weighting in the objective function. A dual-stage voxel division strategy is employed to first broaden the matching scope and then refine the registration precision. Experimental results demonstrate that the proposed method reduces the median translation error by 50% (from 1.0 × 10⁻³ m to 0.5 × 10⁻³ m) and the median rotation error by 50% (from 0.01° to 0.005°), compared to the original 3D-NDT. Additionally, it significantly improves robustness and accuracy, doubling the effective translation range (from 1.2 m to 2.4 m) and increasing the rotation range by 25.6% (from 39° to 49°).