PointTr: Low-Overlap Point Cloud Registration With Transformer

Abstract

Point cloud registration is not only a key step in constructing high-precision 3-D maps, but is also of great significance in fields such as autonomous navigation, environmental perception, and robot operation. However, in outdoor environments, point cloud registration tasks face challenges such as complex terrains and object occlusion, leading to significant overlap and incomplete data. To address these challenges, this article proposes a novel approach that utilizes geometric Transformer for precise point cloud registration. This method employs a learnable geometric position update (GPU) module that automatically learns and captures the geometric structure and features among point cloud data. Additionally, the algorithm introduces a deeper cross-attention (DCA) module that applies deep convolutional operations on the key and value of cross-attention, enabling the attention mechanism to capture both local spatial structures and global contextual information of the point cloud. Through information interaction, the expressiveness and adaptability of the point cloud registration model can be enhanced. Experimental evaluations on multiple benchmark tests, including 3DMatch/3DLoMatch, KITTI, ModelNet/ModelLoNet, and MVP-RG, demonstrate the competitive performance of our algorithm. Compared to the UDPReg and REGTR algorithms, our method demonstrates performance improvements in the RR metric, increasing by 11.3% and 10.8%, respectively, in the low-overlap case.