Improving point cloud registration accuracy under low overlap conditions based on deep learning

Point cloud registration is crucial for analyzing and utilizing point cloud data. However, in specific scenarios, the overlap between the target and source point clouds is relatively low, significantly reducing the success rate of point cloud registration. To address this challenge, we propose modifications to enhance the Representative Overlapping Points Network and achieve notable improvements. In the initial registration stage, we employ a Point Cloud Transformer network with an additional attention mechanism for feature extraction. This novel network architecture enhances our understanding of global feature information. Furthermore, we introduce a new convolutional network design to predict the overlap between two types of point clouds more accurately and utilize a novel loss function for iterative deep learning. Experiments on the ModelNet40 dataset were conducted to assess the efficacy of the proposed method. The registration rate in point cloud registration with low overlap specifically increased by 1.87%, while Error (R) was reduced by 12.75%, and there was a decrease of 5.86% in loss error.