Part Point Cloud Segmentation and High-Precision Localization via Geometric Attention and Edge Convolution

To address issues such as noise interference, occlusion-induced incompleteness, and extreme foreground-background class imbalance in industrial part point clouds—factors that often lead to blurred semantic segmentation boundaries, missed detection of small objects, and reduced 3D localization accuracy—an enhanced PointNet++ network integrating geometric attention and edge convolution was proposed. First, a local geometric refinement module, EdgeConv, was introduced in the encoding stage to explicitly model neighborhood topological relationships and enhance the representation of edge details. Meanwhile, a one-dimensional multi-scale efficient attention module, EMA_1D, was designed to dynamically weight features, suppressing background noise responses and highlighting key structural characteristics. Furthermore, regarding the optimization objective, a hybrid loss function combining Lovász-Softmax and negative log-likelihood was adopted to directly optimize the intersection over union metric, effectively alleviating model bias caused by class imbalance. Experimental results show that a mean intersection over union (mIoU) of 93.13% is achieved by the proposed method on the industrial part segmentation task, while high overall accuracy and per-class accuracy are maintained. Furthermore, in the end-to-end localization pipeline, millimeter-level 3D localization accuracy is achieved by integrating segmentation results, registration, and pose normalization strategies, with an average localization error of 1.24 mm, and the effectiveness and practical potential of the proposed method for high-precision 3D localization under complex industrial point cloud conditions are validated. In this study, effective technical support for solving the 3D perception challenges in scenarios with strong interference and weak features is provided, which is of great significance for enhancing the autonomous operation capability of industrial robots.