Cross-PCR: A Robust Cross-Source Point Cloud Registration Framework

TL;DR

Cross-PCR tackles cross-source point cloud registration by addressing density differences with a density-robust encoder and by transforming difficult feature matching into a two-stage, loosely generated but strictly filtered process. It combines a multi-density fusion strategy, one-to-many sparse matching with spectral consistency, and a prior-guided global dense matching scheme, followed by pose estimation via weighted SVD ICP and a robust hypothesis selection scheme using sparse Chamfer-like distance. The approach achieves state-of-the-art performance on cross-source data, notably boosting registration recall by 57.6 pp and feature matching recall by 63.5 pp on Kinect-LiDAR in 3DCSR, and also delivers top results on 3DMatch while maintaining robustness to downsampling. These results demonstrate significant practical impact for reliable multi-sensor fusion in robotics and mapping tasks, including indoor scenarios involving Kinect and LiDAR sensors.

Abstract

Due to the density inconsistency and distribution difference between cross-source point clouds, previous methods fail in cross-source point cloud registration. We propose a density-robust feature extraction and matching scheme to achieve robust and accurate cross-source registration. To address the density inconsistency between cross-source data, we introduce a density-robust encoder for extracting density-robust features. To tackle the issue of challenging feature matching and few correct correspondences, we adopt a loose-to-strict matching pipeline with a ``loose generation, strict selection'' idea. Under it, we employ a one-to-many strategy to loosely generate initial correspondences. Subsequently, high-quality correspondences are strictly selected to achieve robust registration through sparse matching and dense matching. On the challenging Kinect-LiDAR scene in the cross-source 3DCSR dataset, our method improves feature matching recall by 63.5 percentage points (pp) and registration recall by 57.6 pp. It also achieves the best performance on 3DMatch, while maintaining robustness under diverse downsampling densities.

Figures (3)

• Figure 1: Existing methods often struggle with density differences and difficult matching, leading to limited correct correspondences. In contrast, our Cross-PCR introduces a density-robust feature extractor to extract density-robust features. Then, a loose-to-strict matching strategy, guided by the "loose generation, strict selection" idea, generates reliable correspondences.
• Figure 2: Our Cross-PCR consists of three parts: density-robust feature extraction, loose-to-strict feature matching, and pose estimation. First, in density-robust feature extraction, a pair of cross-source point clouds $\mathbf{P}$ and $\mathbf{Q}$ are fed into a density-robust encoder to obtain the density-robust features $\mathbf{\widetilde{F} }^{\mathcal{{P}}}$ and $\mathbf{\widetilde{F} }^{\mathcal{{Q}}}$. After that, in loose-to-strict feature matching, we propose a two-stage matching, which follows a "loose generation, strict selection" strategy to generate $m$ sets of robust and reliable correspondences $\mathcal{\widetilde{G}}_i^{\prime}$. Finally, in pose estimation, we propose a novel hypothesis selection method to select the best transformation $\mathbf{{T}}^{*}$.
• Figure 3: Correspondences and registration results compared with GeoTrans qin2022geometric