Simulation of Self-Occlusion Virtual Dataset Method for Robust Point Matching Algorithm, With Applications to Positioning of Guide Vanes

The application of point cloud registration technology for workpiece positioning compensation using optical three-dimensional measurement methods has attracted widespread attention in the manufacturing industry, particularly point cloud registration methods integrated with deep learning are booming. Since the training of current deep learning registration methods is often based on public datasets, the performance of point cloud registration of guide vanes depends on the relevance, quality, and quantity of the training dataset, if the training is directly based on the current public dataset used for guide vanes, the accuracy of the registration criteria cannot meet the requirements. Additionally, in real industrial scenarios, manually obtaining the real dataset is time-consuming, labor-intensive, and error-prone. To address these issues, this paper proposes a virtual simulation method based on the CAD model of the workpiece to set up a virtual camera so that a large number of near-real datasets can be generated quickly. The method can simulate the incomplete vane point cloud obtained by real shooting due to self-occlusion by setting multi-angle virtual cameras on the hemispherical surface wrapped in the CAD model. The experimental results show that the combination of the deep learning registration method and the virtual dataset method in this paper can improve the accuracy, efficiency, and stability of the deep learning registration for workpiece positioning compensation, which has a good prospect for practical application.