RoCNet ‐Pro: Nuclear Reactor Pressure Vessel Point Cloud Registration Based on Multi‐Scale Curvature Uncertainty Quantitative Feature Descriptor

Abstract

ABSTRACT To address the challenge of point cloud registration for highly symmetrical structures such as nuclear reactor pressure vessels, this paper proposes RoCNet‐pro: a registration method based on multi‐scale curvature uncertainty quantification features. This method breaks through the key bottlenecks, such as feature blurring and matching ambiguity caused by high symmetry, and achieves a significant improvement in the accuracy of symmetric structure registration. The main contributions of RoCNet‐pro include: (1) A multi‐scale curvature feature extractor combined with uncertainty quantification is developed. The geometric characteristics are used to enhance the robustness and adaptability of features, and the registration accuracy is significantly improved. (2) In this paper, the edge slack block module is designed. By introducing the edge relaxation mechanism to the similarity matrix, the discrimination problem caused by local feature blurring in the symmetric structure is effectively solved. (3) The deformable self‐attention mechanism and shared module are introduced to effectively reduce the computational complexity and enhance the generalization ability and training efficiency of the network.