Overlapping Prediction Network for Local-to-Global Point Cloud Registration

Local-to-global point cloud registration is a key technology in computer-assisted orthopedic surgery. However, due to the little overlapping between local and global point clouds, there simply are not enough reliably corresponded point pairs in the course of the registration to make registration algorithm design and realization easy. To solve this problem, we propose the overlapping prediction network based on the attention mechanism to better extract point cloud features for improved prediction of overlapping regions between the two point clouds as an important basis for further registration. First, an alignment module utilizing a simple encoder was proposed for moving the source point cloud very close to the target point cloud to facilitate further fine registration. Then, we designed an overlapping point prediction module that detects overlapping regions based on point similarity to damp the influence of non-overlapping points. Feature extraction is the core of the model, where the attention mechanism is utilized to make it a closer representation of the topological structure of the point cloud, leading to efficient global feature extraction. Experiments show that with the introduction of our overlapping prediction into existing registration algorithms, the performance of point cloud registration dramatically improved. Compared to existing overlapping detection mechanisms, ours outshines others across the metrics of accuracy, precision, recall, and F1 score, thus evidencing its efficacy.