HD-BTC: Hierarchical Density-Enhanced Binary and Triangle Combined Descriptor With Plane Assistance for Air–Ground Loop Closing

Abstract

Achieving reliable loop closing between aerial and ground robots is essential for heterogeneous multi-robot SLAM. However, existing methods often rely on homogeneous sensor configurations or are restricted to specific environments (either indoor or outdoor), limiting their generalizability across varying platforms and scenes. To address these challenges, this paper proposes a hierarchical density-enhanced binary and triangle combined descriptor with plane assistance for air-ground loop closing. First, an adaptive multi-frame fusion strategy is developed by incorporating sensor height, installation angle, and horizontal velocity, enabling rapid adaptation to diverse LiDARs on heterogeneous robots. Second, to address the indoor structural ambiguity, the plane-assisted ceiling-ground segmentation method is proposed. It can select ground while filtering out ceiling interference. Furthermore, to cope with air-ground viewpoint differences in outdoor scenes, we leverage a hierarchical density-enhanced BTC descriptor, and incorporate the incremental keypoints extraction and multi-layer density-guided loop detection for acceleration. Extensive benchmarks against state-of-the-art methods, such as GAPR, EHPR, and UniLGL, demonstrate that our method achieves superior performance on both indoor and outdoor air-ground datasets, while maintaining competitive generalization on cross-LiDAR multi-session datasets, with only a modest 2.9 ms runtime increase over BTC. Note to Practitioners—This work is motivated by the practical difficulty of achieving reliable loop closing in collaborative system composed of aerial and ground robots equipped with different LiDARs. In many real-world applications, such as infrastructure inspection, underground facility monitoring, and emergency response, heterogeneous robots are required to operate across both indoor and outdoor environments. In practice, the reliability of existing methods is often degraded by LiDAR differences, indoor structural interference such as ceilings, and outdoor large air-ground viewpoint discrepancies. This paper develops a LiDAR-based loop closing framework that explicitly accounts for sensor heterogeneity and environment diversity. Instead of assuming fixed sensor configurations or uniform operating conditions, the method adaptively constructs consistent local submaps. It further improves robustness in indoor environments by mitigating ceiling ambiguities, and construct HD-BTC descriptor for solving the outdoor air-ground viewpoint differences. The proposed method has been evaluated via public and collected datasets. Future work will focus on improving robustness in highly dynamic or open spaces where geometric features are sparse.