A Novel Point Cloud Segmentation Method for Accurate Surface Partitioning Based on Feature Boundaries in the Machining of Industrial Components

The surfaces of industrial components used in aerospace fields typically exhibit diverse feature boundaries. These features divide the surface into multiple sub-regions, and effective machining for such components therefore depends heavily on region-based partitioning techniques. Employing measured surface points for regional segmentation and feature extraction has proven to be a practical and effective approach. However, due to the multi-scale and smoothly varying nature of these surface features, existing point cloud segmentation methods often struggle to deliver satisfactory results in such contexts. To address this limitation, this paper proposes a feature clustering-based segmentation method specifically designed for such scenarios. The robust identification advantage of smooth features in manifold space is first leveraged through tensor voting theory, with a multiscale voting scheme being established to extract the fine feature boundaries of components. Furthermore, a nearest neighbor adjacency topology based on Delaunay triangulation is constructed in the neighborhood of these boundaries, derived from the properties of Voronoi diagram space partitioning. This facilitates the perception of behavior crossing feature boundaries during the Euclidean distance clustering process and ensures the effective clustering of enclosed feature regions. The proposed method has been experimentally validated on four industrial component datasets with different types of features. Both qualitative and quantitative analyses of the segmentation results demonstrate that the method achieves approximately 90% in precision and recall. Compared with several classical and state-of-the-art approaches, it consistently achieves top performance across various evaluation metrics. The experimental results confirm that the proposed method offers significant improvements in the accuracy and effectiveness of component-level point cloud feature recognition and region segmentation.