A dynamic learning framework integrating attention mechanism for point cloud registration

To improve the low accuracy problem of existing point cloud registration algorithms attributed to deficient point cloud geometric features, we proposed a new point cloud registration network inspired by dynamic feature extraction and the graph attention mechanism. The model uses the dynamic graph edge convolution neural network to characterize the multi-level semantics of the point cloud at first, then uses a feature fusion module based on attention mechanism to fuse the representation information, and finally uses the singular value decomposition (SVD) method to generate the transformation matrix. The experimental verification was carried out on the ModelNet40, ShapeNet Part datasets, and the local industrial part dataset. Experiment results show that our model gets competitive registration performance compared with other advanced models on three datasets. When tested on the untrained data class and the noisy circumstances, our model gets lower average registration errors than compared models. It shows that our framework has not only the characteristics of high registration accuracy and generalization ability but also strong robustness.