E2MNet: an end-to-end large-scale point cloud registration network based on Mamba

When handling large-scale outdoor point cloud data, the performance of deep learning–based point cloud registration networks is significantly limited mainly due to factors such as massive data volume, dynamic moving objects, and extensive environmental noise. To address this, we propose E2MNet, an end-to-end large-scale point cloud registration network based on the Mamba framework. The network is capable of efficiently processing point clouds with more than 180,000 points per frame on a single NVIDIA GeForce RTX 2060 GPU (with 8 GB memory). E2MNet employs a hierarchical design. During the feature extraction stage, an improved state space duality module combined with depthwise separable convolutions not only comprehensively captures both local and global features of point clouds but also effectively mitigates interference from noise. In the fusion and registration phase, a spatio-temporal fusion module based on the state space model is designed to jointly model the spatio-temporal relationship between two frames of point clouds, efficiently handling the changes caused by dynamic objects. Through extensive experiments on the KITTI and NuScenes datasets, E2MNet achieves the highest registration recall rates of 99.8% and 99.9% with the fastest processing speeds while also minimizing relative translation errors, thoroughly validating its effectiveness in large-scale outdoor point cloud registration tasks and offering an idea and approach for research in this field.