Abstract
Point cloud registration is a fundamental task in autonomous driving and 3D mapping, yet large-scale LiDAR data often presents significant challenges due to high outlier ratios and computational complexity. This paper introduces a robust, decoupled registration framework that leverages plane primitives instead of individual points. By combining a "guaranteed" outlier removal strategy for rotation with a component-wise estimation for translation, the proposed method achieves state-of-the-art accuracy and efficiency, significantly outperforming both classical and deep learning-based methods on massive outdoor datasets.
Introduction
Traditional point-based registration methods, such as Iterative Closest Point (ICP), frequently struggle with local optima in the absence of precise initial alignments. Global alternatives often rely on RANSAC, which suffers from exponential slowdowns as outlier counts increase. Furthermore, point-based descriptors are highly sensitive to variations in point density and viewpoint common in LiDAR scans.
This research addresses these bottlenecks by shifting the focus from points to planes. Planes are more stable, information-dense geometric features in human-made environments. By decoupling the 6-degree-of-freedom (DoF) transformation into separate rotation and translation stages, the optimization problem is simplified, allowing for more reliable and efficient global alignment.
Methodology
The proposed framework follows a structured pipeline that moves from raw point clouds to a final aligned result through three primary stages: preprocessing, rotation estimation, and translation estimation.
1. Rotation Estimation with Guaranteed Outlier Removal
The rotation estimation stage is designed to be resilient against incorrect plane correspondences through a multi-tiered filtering approach:
- Coarse Filtering: Uses rotation-invariant constraints, such as the angles between plane normals, to rapidly discard obvious mismatches.
- Refined Filtering: Employs rigorous computational geometric consistency checks. For each candidate match, the system calculates "rotation bounds"—the specific range of angles that could potentially align the planes within a given error threshold. If no such angle exists, the match is definitively discarded.
- Interval Stabbing: This technique identifies the rotation angle that satisfies the maximum number of candidate matches simultaneously, ensuring a globally optimal rotation estimate.
2. Component-Wise Translation Estimation (CWTE)
Following rotation, the 3-DoF translation is resolved. Instead of searching the entire 3D space, the method breaks the translation down into 1-DoF components. Since parallel planes constrain translation in only one dimension, correspondences are grouped into "component sets." Each component is estimated globally, and the results are combined to resolve the full 3D translation vector.
Experimental Results
The method was evaluated on several large-scale LiDAR benchmarks, including ETH3D, ETH-TLS, and WHU-TLS.
- Accuracy: On the WHU-TLS dataset, the method achieved a mean rotation error of 0.752° and a translation error of 0.509m. This represents a massive improvement over established methods like Teaser++ (105.4° / 78.59m) and learning-based approaches like RoReg (24.86° / 52.92m).
- Efficiency: The core estimation phase is remarkably fast, typically completing in less than 0.3 seconds.
- Data Parsimony: The use of plane primitives allows for extreme data compression. The method requires only 480 bytes per cloud, compared to approximately 1 million bytes for descriptor-heavy methods like Teaser++, making it ideal for bandwidth-constrained real-time applications.
Conclusion
By leveraging the geometric stability of planes and employing a decoupled, guaranteed outlier removal strategy, this framework provides a highly accurate and efficient solution for large-scale LiDAR registration. While currently optimized for structured environments where planes are prevalent, future work aims to extend this approach to multi-view registration and integrate it with deep learning for even more robust feature extraction in unstructured natural scenes.
Summary updated