GPAT: Enhancing point cloud registration with local-global Integration and Geo-positional awareness

Point cloud registration, essential for applications like autonomous driving and 3D reconstruction, remains challenging in scenarios with low overlap and incomplete data. Although learning-based methods have shown promise, existing keypoint-based and coarse-to-fine approaches can struggle with ambiguous matching due to similar local patches in non-overlapping regions, particularly under low overlap conditions. This study introduces GPAT(Geometric-Positional Attention and Local-Global Transformer Networks), a novel hierarchical deep learning framework designed to address these limitations by effectively integrating local geometric and positional information with global structural awareness. By leveraging a novel positional encoding and attention mechanisms, GPAT captures the intricate interplay between geometric structures and spatial configurations, enriching the distinctiveness of local features. Simultaneously, it incorporates global geometric consistency to establish robust correspondences, effectively bridging the gap between local and global contexts. The framework further refines these correspondences into dense point matches, ensuring precise registration even under challenging conditions. Extensive experiments on benchmark datasets, including 3DMatch, ModelNet, and KITTI, demonstrate GPAT's competitive performance, especially in low-overlap situations, demonstrating its capability to extract discriminative features and achieve accurate transformations with high robustness.