Point cloud registration method integrating global context awareness and cross-point-cloud interaction

Point cloud registration is a fundamental technique in point cloud data processing and finds widespread applications in fields such as robotics and industrial inspection. However, the lack of global contextual information in point cloud features and the presence of feature mismatches between point pairs often result in inaccurate rigid transformation matrix estimation, leading to low registration accuracy. To address these challenges, this paper proposes a point cloud registration method integrating global context awareness and cross-point-cloud interactions. The method comprises two primary modules: a global-local feature extraction module and an attention-based feature fusion and matching module. The global-local feature extraction module extracts local geometric features from the point cloud while incorporating global contextual information to enhance the global consistency of point cloud features. The attention-based feature fusion and matching module employs a cross-attention mechanism to integrate feature information from two point clouds and estimates the confidence of point-pair correspondences. Then, a weighted singular value decomposition algorithm was implemented to improve the estimation accuracy of the rigid transformation matrix. Experimental results show that the proposed method achieves a feature matching recall (FMR) of 97.9% on the 3DMatch dataset, surpassing existing methods.