3D Pairwise Point Cloud Registration Network for Power Equipment Based on Multi-scale Features

The diverse range of power equipment, each with its unique shape, poses a challenge for precise registration. This paper advances a 3D paired point cloud alignment model that is founded on multi-scale feature extraction, attempting to alleviate the negative impact of outliers from two aspects. Initially, a multi-scale feature extraction model is formulated to enhance structural information across various levels, thereby enhancing the discriminative capabilities of point features. Secondly, the transformation matrix is utilized to extract the internal points of the source and target point cloud. A constraint function is then incorporated to steer the registration process, enhancing the inductive bias of the model and augmenting its generalization capabilities. Finally, the applicability of this method is verified on power automation equipment of various shapes. Experimental conclusions indicate that the method which described in this paper attains an MAE value of 0.31 on the ModelNet40 dataset, while also effectively registering previously unseen power equipment, showing good generalization ability and practical application potential.