SegCert‐PCR: Ground Segmentation Based Certifiable Point Cloud Registration in Unstructured Field Environments

ABSTRACT Point cloud registration is fundamental to 3D perception for robot navigation, autonomous driving, SLAM, and map construction. Although indoor registration has advanced substantially, complex unstructured outdoor environments remain challenging because of weak geometric features, limited structural constraints, large datasets, dynamic interference, and large interframe motion. We propose SegCert‐PCR, a robust and certifiable registration framework that combines adaptive ground segmentation with certifiable global registration. A terrain‐aware segmentation module removes dominant ground points while retaining discriminative nonground structures. FPFH descriptors and bidirectional consistency matching then select high‐confidence point correspondences. Registration is performed in two stages. First, a graph‐theoretic maximum‐consensus optimizer with TEASER++‐style certifiable robustness rejects outliers and estimates a reliable global transformation. Second, point‐to‐plane ICP refines this estimate for high‐precision local alignment. Experiments on representative RELLIS‐3D off‐road scenes demonstrate improved robustness and stability relative to state‐of‐the‐art methods across large frame intervals. Rotation errors range from to , and translation errors range from 0.0343 to 0.0747 m. These results show that SegCert‐PCR provides stable alignment under sparse geometry, ground redundancy, noise, dynamic disturbances, and substantial viewpoint changes, making it suitable for field‐robotics perception in unstructured environments.