Swift and Accurate Point Cloud Registration using SGUformer

Point cloud registration aims to find a rigid transformation that aligns two point clouds. Applications such as augmented reality (AR) and robot navigation often require real-time performance for point cloud registration algorithms. In this paper, we propose SGUformer, a novel point cloud registration method that achieves fast alignment by redesigning the feature extraction pipeline and employing a lightweight global feature extraction framework. The gating mechanism is utilized and local coordinates are embedded to enhance the representation of point-level features. To facilitate the extraction of global features, a transformer with 3D rotary position embedding is implemented, circumventing the need to compute relative position information, thereby improving computational efficiency. Furthermore, a part attention mechanism is designed to tackle outlier pollution issues. In the final registration stage, the registration results obtained from each patch pair, weighted by their respective confidence scores, are combined to vote and acquire a more robust final result. In the conducted experiment, the superior quality of features derived from the novel structure’s feature extractor enabled our method to attain a better Feature Matching Recall (FMR) in comparison to existing leading methodologies. Moreover, the implementation of the proposed registration method resulted in the highest recorded registration success rate, exceeding the second-best method by 0.8%. Additionally, our approach demonstrated a remarkable efficiency, being 26% faster than the alternative methods.