Research on Point Cloud Registration and Stitching Fusion Algorithm Based on GCN-PRFNet

Point-cloud registration and stitching are important topics in the field of robot navigation and 3D reconstruction, e.g., the accuracy of point cloud registration and stitching in robot navigation directly affects the accuracy of map construction. Many researchers have proposed various algorithms for deep learning-based point cloud registration and stitching methods with good performance, and although there are end-to-end methods that have made progress, they still have limitations in local feature fusion efficiency and geometric detail retention. To address this issue, a fusion algorithm for registration and stitching based on a GCN-PRFNet point cloud is proposed. The network has a feature extraction module, a point cloud registration module, and a point cloud splicing and fusion module. GCN-PRFNet can efficiently handle the task of point cloud registration and splicing and fusion in partially overlapping regions and is robust to noise. The model is trained on the ModelNet40 dataset, and its registration and splicing accuracies are improved by 53.9%, 20.1%, 8.3%, 12.2%, 6.1% and 1.8% when compared with the traditional iterative closest point and learning-based PointNetLK, DGCNN, RPM-Net, DCP, and PointViG methods. This indicates that the constructed model is effective in point cloud registration and splicing. Meanwhile, point-cloud registration and splicing tests were performed on five self-constructed artefact datasets, and their registration and splicing accuracies were over 90%, indicating that the constructed end-to-end point-cloud registration and splicing model is considerably effective in real-world application scenarios.