DRIR-Net: Dual-branch rotation invariant and robust network for 3D place recognition

Place recognition using 3D point clouds (LiDAR) is crucial in autonomous driving, especially for simultaneous localization and mapping (SLAM) in challenging outdoor environments. Nowadays learning-based methods face significant challenges, such as substantial preprocessing time, difficulty in handling viewpoint rotation, and inability to resist noise interference. To address these issues, we propose a Dual-Branch Rotation Invariant and Robust Network (DRIR-Net) for 3D place recognition. In DRIR-Net, we leverage the characteristics of LiDAR scanning to segment the point cloud and employ a subregion downsampling method to reduce preprocessing time significantly. Utilizing this subregion structure, we design a local position encoding and feature extraction module to achieve robustness to data noise. Furthermore, we propose a global feature augmentation module for subregions and generate rotation-invariant global descriptors combined with symmetry functions. Additionally, we improve place recognition performance by fusing point and voxel features with subregion cross attention. Through extensive experiments, our method outperforms state-of-the-art approaches on the USyd, Oxford RobotCar, and KITTI datasets. Importantly, it effectively handles scenes with noise from rotated viewpoints and generalizes well across different environments.