TopSPR-Net: Topology Aware Segment-Level Point Cloud Learning Descriptors for Three-Dimensional Place Recognition in Large-Scale Environments

Ensuring precise data association is pivotal for reliable perception and localization of agents navigating outdoor large-scale environments. However, existing methods that rely on point cloud proxy representations (e.g., voxels, range image views, and bird's-eye views) still face limitations when it comes to challenges such as viewpoint and appearance changes. To this end, we propose a topology-aware place recognition method called TopSPR-Net, which explores segment-level point clouds to extract more unique and discriminative descriptors. Specifically, in terms of point cloud preprocessing, an efficient and effective clustering method is implemented to achieve distinct segmentation. In terms of segment-level feature embedding, the unified advantages of convolutional neural networks and pruning Transformer are adeptly explored to enhance feature saliency and discrimination by selectively emphasizing long-range dependencies and interactions. Extensive evaluations on KITTI, KITTI-360, and real vehicle datasets demonstrate that our customized TopSPR-Net achieves an encouraging state-of-the-art, with an average F-1 max score of 0.98 compared to other representative baselines.