Abstract: The features extracted by the existing point cloud registration algorithms are not so rich, which makes it difficult to further improve the accuracy of the registration. To address this problem, a deep learning-based multi-scale feature fusion point cloud registration algorithm is proposed. EdgeConv is employed to extract multiple features of different scales through the algorithm at first, which can maintain the local geometric structure characteristics. Then Non-linear Polarized Self-attention is introduced to filter its output features, and thus the effectiveness of feature information is improved. And later the above multi-scale features are fused and EdgeConv is employed again to extract their features, thereby improving the expression ability of the features. In the rigid pose estimation stage, Lie algebra is used to process the rotational transformation to fully exploit the transformation information of the point cloud. According to the changes of the extracted point cloud features during the registration process, the weight values of the components of the loss function are dynamically adjusted to evaluate the prediction results of the model more accurately. Tested on the ModelNet40 dataset, when the sample types of the train and test sets are the same, the rotation error of the proposed algorithm is 1.826 7 and the displacement error is 0.001 0, and when the sample types of the train and test sets are not the same (experiments on generalization), the rotation error of the proposed algorithm is 2.979 4 and the displacement error is 0.001 0. The experimental results show that the registration accuracy of the proposed algorithm has improved compared with the current mainstream algorithms, and it exhibits good generalization performance.

Key words: deep learning, point cloud registration, feature extraction, rigid object, pose estimation, Lie algebra