Local–global transformer-based point cloud segmentation network for workpiece surface defect grinding

Point cloud is extremely useful in high-precision intelligent robotic grinding due to their abundant spatial information. Accurate point cloud segmentation can extract surface defects of workpieces to provide 3D data for robotic grinding. Recently, with the development of computer vision, Transformer-based methods have been widely applied in point cloud segmentation, while most of them only capture features based on global or local attention, resulting in a lack of comprehensive information. In this paper, we propose a novel segmentation network, Local and Global Feature Transformer Network (LGFT-Net), to combine local features and global features based on Transformer module. Specifically, a local feature extraction block with feature embedding module and Local Transformer module is designed to embed features into a higher space and learn feature correlations in local regions, and a global feature extraction block with Global Transformer module is used to obtain global contextual information. We perform extensive experiments on ShapeNet dataset and workpiece point cloud dataset. The results show that: (1) the LGFT-Net achieves segmentation performance comparable to state-of-the-art methods in part segmentation and (2) this model is feasible to segment surface bump defects of workpieces to provide 3D data for robotic grinding.