Low Overlap Point Cloud Registration Method Based on Deep Position-aware Transformer

Abstract

In response to the issues such as neglecting the fusion of local geometric embeddings in the feature extraction stage,weak correlation in position-aware information between low overlap point cloud pairs in the feature interaction stage,making it difficult to extract more expressive features and deviation in the transformation solved due to some outlier correspondence in the correspondence generation stage,in this paper,a 3D point cloud low overlap registration method based on deep position-aware Transformer(DeepPAT) is proposed,which follows the local to global matching mechanism.A local feature extraction network based on local geometry information is proposed to extract multi-level features from point cloud.Then,a deep position-aware Transformer(DPAT) module is designed to extract the relevant features and overlap information between low overlap point cloud pairs by learning the geometry and spatial position information of the point cloud itself and across frames,so as to carry out low overlap point cloud matching.Finally,a maximal cliques algorithm adjusted by the feature similarity is designed to reduce the position ambiguity caused by the length consistency and eliminate the outlier correspondences.It can be used as a plug-and-play robust estimator to replace traditional robust estimators such as RANSAC and is fully implemented by Pytorch.Evaluating on the synthetic ModelNet dataset and indoor 3DMatch dataset,the experimental results show that DeepPAT reduces the rotation and translation root mean square error to 3.994 and 0.005 on ModelNet datasets,respectively,and DeepPAT outperformed existing methods by at least 4.3 percentage points and 3.6 percentage points in term of registration recall on 3DMatch and 3DLoMatch benchmarks,respectively.