Multi-level point cloud transform regression via overlapping region prediction

Accurate and efficient point cloud registration is an important application of artificial intelligence for precise navigation and reliable decision-making in autonomous driving. However, the down-sampling of point clouds introduces interference from non-overlapping regions, resulting in the loss of important geometric information and a reduction in registration accuracy. Therefore, this study proposes an overlap-aware multi-level transformation regression network (OMTRNet), a deep-learning model that couples overlap prediction with frequency-domain and spatial feature interaction. Through a hierarchical feature extraction module designed with pixel grid mapping, we leverage the structural characteristics of pixel grids to efficiently extract point cloud features across multiple scales. By using a convolutional neural network-based overlap region prediction module, the overlap between the source and target point clouds at different scales is predicted. The overlapping region of the point cloud features is extracted from the predicted overlap, allowing for further feature extraction. Additionally, a frequency-domain feature interaction module is introduced to combine low-frequency global structural information with high-frequency local details as complementary cues for transformation estimation. Experimental results on the KITTI and NuScenes datasets validate the effectiveness of OMTRNet for large-scale partially overlapping point cloud registration. On the KITTI dataset, OMTRNet achieves an RTE of 0.2675 m, an RRE of 0.5342 deg, and a registration recall of 92.10%. The ablation results show that the DCT-based frequency-spatial interaction achieves competitive registration performance among the compared feature-interaction strategies.