关键词: 点云分割, 机械零件, 三维点云, 语义分割

Abstract: In modern industrial automation,point cloud segmentation is a key technology for understanding complex scenes.However,in complex scenarios involving stacked mechanical parts,existing segmentation methods often struggle with challenges such as uneven point cloud density,noise interference,and severe occlusion,resulting in suboptimal segmentation accuracy.To address these challenges,an innovative backbone network that combined Transformer and Mamba named TMPNet was proposed for point cloud segmentation of stacked mechanical parts.TMPNet employed an adaptive curve aggregation sorting strategy that dynamically adjusted the point cloud sorting order,effectively preserving both global and local features of the point cloud and enhancing the model's sensitivity to different shapes and structures.Additionally,the use of multidimensional geometric positional encoding enabled a more detailed representation of the geometric information within the point cloud,further improving the model's segmentation accuracy and stability in complex stacked scenarios.Experimental results demonstrated that TMPNet significantly improved mIoU and mAcc on our custom-built point cloud segmentation dataset of stacked mechanical parts and showed strong generalization capabilities on the public S3DIS dataset,which provided a more reliable foundation for robotic grasping and subsequent point cloud completion tasks.

Key words: point cloud segmentation, mechanical parts, 3D point cloud, semantic segmentation