基于图注意力的铸件飞边点云识别算法

Point Cloud Recognition Algorithm for Casting Flash Based on Deep Learning

针对铸件飞边打磨加工过程中,零件需要定制特定安装夹具和机器人打磨需提前离线编程,但仍难适应铸件飞边个性化特点,提出一种面向三维视觉引导机器人打磨的铸件飞边点云识别算法,在线分割出飞边点云,为后续打磨加工飞边提供精准数据支撑。该算法基于图注意力(GAC-Net)模型,提出以对称函数激活和最远点采样算法融合的方式进行降采样,为克服点云的无序性,融合排列不变卷积信息,为扩大网络视野,在上采样过程中融合全局信息。实验结果表明,上述算法m-IOU达到了97.80%,比GAC-Net算法高出5.34%,说明了所提算法的使用性。通过消融实验,论证了提出的特征融合、降采样和上采样方法的有效性。

In the process of casting flash grinding, parts need to customize specific installation fixtures and robot grinding needs to be programmed offline in advance, but it is still difficult to adapt to the personalized characteristics of casting flash. A recognition algorithm of casting flash point cloud oriented to 3D vision guided robot grinding is proposed to segment the flash point cloud on the line, providing accurate data support for subsequent grinding flash. Based on the graph attention(GAC Net) model, the algorithm proposes a method of combining symmetric function activation with the farthest point sampling algorithm to reduce sampling. In order to overcome the disorder of point clouds, the information of permutation invariant convolution is fused;In order to expand the network view, the global information is fused in the up sampling process. The experimental results show that the m-IOU of the above algorithm reaches 97.80%,5.34% higher than that of the GAC-Net algorithm, which shows the usability of the proposed algorithm. Through ablation experiments, the effectiveness of the proposed feature fusion, downsampling and upsampling methods is demonstrated.