Towards High-Accuracy Point Cloud Registration with Channel Self-attention and Angle Invariance

Point cloud registration, which plays a vital role in various applications, remains a very challenging task, especially in scenarios with low overlap between source and target 3D point clouds. This paper proposes a high-accuracy and robust point cloud registration method by designing a fused geometric feature extraction module and a two-stage outlier pruning strategy. In the fused geometric feature extraction module, we introduce a non-local feature aggregation block that adopts a channel self-attention mechanism, named the CHNonlocal block, to enhance geometric features extracted from point-to-point correspondences using the known SCNonlocal block. The two-stage outlier pruning strategy further refines outlier removal by leveraging angle invariance to obtain consensus sets targeting point-to-point correspondences, initially identified by pruning outliers based on feature similarity. Similar to PointDSC, candidate pose transformation parameters can be calculated from these consensus sets, and the best hypothesis can be selected as the final registration result. Experiments on various point cloud datasets demonstrate that our method achieves higher registration accuracy and better generalization ability compared to mainstream point cloud registration algorithms, especially in low-overlap scenarios.