Abstract: The accuracy of digital process simulation model is the core of body welding process development,and the segmentation and identification of process equipment based on point cloud in the welding scene is the key to realize the consistency between physical operation environment and simulation environment.For the problems of uneven point cloud density,large local feature differences and dependence on labeled samples in point cloud segmentation of welding assembly site,a semi-supervised point cloud semantic segmentation method based on generative adversarial network under the condition of few samples was proposed,and the segmentation accuracy by fusing the labeled and unlabeled point cloud data was improved.The RandLA-Net was improved by adopting the farthest point sampling and adjusting the coding and decoding structure to enhance the complex feature learning ability.The adversarial structure and self-training mechanism were introduced to utilize the unlabeled sample information fully.Furthermore,the segmentation performance was improved by introducing smoothness constraints and selecting highly reliable pseudo-labels to reduce the introduction of mislabeled information.The experimental validation was performed on the self-generated welding station point cloud.Results showed that on the welding station dataset,the proposed method achieved a comparable segmentation performance to that of the fully supervised method,and the segmentation accuracy was improved by 8.40% compared with the semi-supervised point cloud segmentation method.

Key words: welding scene, point cloud semantic segmentation, semi-supervised, generative adversarial network

摘要： 数字化工艺仿真模型的准确性是车身焊装工艺开发的核心,焊装场景下基于点云的工艺装备等分割识别是实现物理作业环境与仿真环境虚实一致性的关键,针对焊装场景点云分割中点云密度不均衡、局部特征差异大、依赖标记样本等问题,提出一种少样本条件下基于生成对抗网络的半监督点云语义分割方法,通过融合使用标记和无标记点云数据来提升分割精度。通过改进RandLA-Net,采用最远点采样并调整编码解码结构以增强复杂特征学习能力;引入对抗结构与自训练机制,充分利用无标记样本信息;通过引入平滑性约束,选择高可靠性伪标签,降低引入错误标签的概率。最后,在自采的焊装工位点云等数据集上开展对比实验验证了所提方法的优越性。

关键词: 焊装工位, 点云语义分割, 半监督, 生成对抗网络