Lightweight and rotation-invariant place recognition network for large-scale raw point clouds

LiDAR-based place recognition is crucial for Simultaneous Localization and Mapping (SLAM) and autonomous driving. However, existing methods face challenges in achieving rotation invariance and are hindered by a large number of parameters and complex data preprocessing steps, such as ground point removal and extensive downsampling. These limitations hinder their practical implementation in complex real-world scenarios. To address these challenges, we propose LR-Net, a lightweight and rotation-invariant place recognition network designed to handle large-scale raw point clouds. LR-Net tackles rotation invariance by efficiently capturing sparse local neighborhood information and generating rotation-invariant features through the analysis of spatial distribution and positional information within the local region. Furthermore, we enhance the network's feature perception by incorporating residual MLP structures to encode multi-level local region features, and integrating dualattentional weighting blocks to emphasize important regions and features. Lastly, the Generalized-Mean pooling structure is utilized to aggregate the global descriptor, enabling efficient point cloud retrieval. Meanwhile, we introduce PRmate, an independent sub-network that effectively optimizes retrieval results by analyzing the descriptor similarity and location distribution of the top 25 matching candidates from the place recognition network. We extensively evaluated the proposed methods on four large-scale datasets, including both standardized point cloud scenes (Oxford RobotCar dataset and NUS In-house dataset) and raw point cloud scenes (MulRan dataset and Kitti odometry dataset). The experimental results demonstrate that LR-Net effectively attains rotation-invariance in place recognition and achieves state-of-the-art accuracy with a lightweight model size of 0.4 M parameters. Additionally, LR-Net eliminates the need for data preprocessing, as it can directly process raw point cloud data for stable place recognition and exhibits robustness against variations in point density and noise intensity, enhancing its suitability for challenging scenarios. Moreover, PRmate is compatible with all existing point cloud-based place recognition methods and improves their performance. These advantages make LR-Net and PRmate highly suitable for robotic applications that rely on LiDAR-based place recognition. Our code is publicly available on the project website (https://github.com/zhzhang023/LR-Net).