An efficient point cloud registration method based on deep learning framework

Abstract

This paper presents an efficient deep learning framework for point cloud registration. Departing from traditional iterative optimization approaches, our method reformulates registration as a regression task to directly predict alignment parameters. The architecture integrates three core components: a point cloud feature extraction network utilizing DGCNN to capture local and global features, a Transformer-based attention network to adjust feature importance adaptively and integrate structural knowledge from different point clouds, and a rigid transformation solution layer to derive the rotation matrix and translation vector. The methodological breakthrough lies in the synergistic integration of these components, enabling direct prediction of registration parameters through learned feature correlation, while maintaining mathematical rigor in transformation estimation. Comprehensive evaluations on the ModelNet40 benchmark demonstrate the framework's high performance, particularly showing remarkable robustness against noise contamination. Quantitative results reveal significant improvements in both computational efficiency and registration accuracy, establishing new state-of-the-art performance for learning-based registration approaches.