SDA-Reg: large-scale dynamic scene point cloud registration with semantic dual-stream attention

Abstract

Point cloud registration is a fundamental task in 3D vision, however, large-scale outdoor LiDAR point clouds, characterized by their immense size and structural complexity, present significant challenges in highly dynamic environments. Existing methods often employ semantic segmentation as preprocessing, offering the potential to incorporate semantic information to enhance registration robustness. This paper introduces SDA-Reg, a registration network based on semantically enhanced features and a dual-stream attention architecture. Its core contributions include: (1) Introducing a semantic consistency constraint module within the attention mechanism to strengthen intra-class correlations and suppress inter-class misalignments; (2) Designing a gated dynamic suppression(GDS) module to adaptively suppress noise from dynamic objects while preserving pseudo-static structures; (3) Deep integration of semantic information into feature extraction and matching processes to achieve semantically guided registration. On the KITTI dataset, SDA-Reg achieves significant performance improvements with a registration recall rate of 99.92%. On the dynamically complex KITTI 08 sequences, it outperforms baseline methods by 1.43%, demonstrating robust accuracy in dynamic environments.