LiDAR-based place recognition is critical for long-term autonomous navigation in Global Navigation Satellite System (GNSS)-denied environments, yet existing methods struggle to balance accuracy and efficiency under substantial yaw rotations. This paper proposes a robust framework based on a multi-channel polar bird’s-eye-view (BEV) representation. Under yaw-dominated revisits, the polar BEV image transforms yaw rotation into cyclic column shifts, providing a useful structural prior for rotation-equivariant feature extraction. Raw point clouds are projected onto polar BEV grids encoding density, height, and intensity. A rotation-equivariant feature extractor comprising a Radial Compression Module and a rotation-equivariant Transformer module captures long-range azimuthal dependencies via Conditional Positional Encoding and Circular Relative-Position Bias. The equivariant features are aggregated by NetVLAD into a compact global descriptor, trained end-to-end with a hard-example mining triplet loss. Extensive experiments on the public KITTI and NCLT datasets, as well as our self-constructed LiDAR Place Recognition Revisit (LPRR) dataset, demonstrate competitive performance on KITTI and superior performance on NCLT and LPRR among the compared methods. The proposed framework achieves a favorable trade-off between performance and computational cost, and shows promising cross-dataset generalization on the evaluated NCLT and LPRR datasets without fine-tuning.
Continual LiDAR place recognition requires a descriptor model to learn new environments while retaining retrieval accuracy in previously learned environments under limited access to historical data. Replaying historical retrieval relations and distilling descriptors impose different optimization demands that fixed loss coefficients cannot readily accommodate. We propose adaptive gradient-balanced boundary memory (AGBBM), a framework that operates on global descriptors without changing the backbone architecture. Boundary-aware triplet memory retains complete anchor–positive–negative relations and their historical distance statistics under an environment-balanced triplet budget. Reliability-guided decoupled distillation separates supervised replay, one-sided positive-distance preservation, and reliability-weighted historical-boundary preservation. A two-level optimization mechanism normalizes active historical losses using exponential moving averages and calibrates same-view descriptor distillation using the ratio of current-learning and distillation gradient norms at a shared descriptor parameter, with configurable attenuation under gradient conflict. We evaluate AGBBM in the stage-wise sequence Oxford → MulRan DCC → MulRan Riverside → In-house using PointNetVLAD, MinkLoc3D, LoGG3D-Net with max pooling, and CASSPR. The final mean Recall@1 values are 61.42%, 79.30%, 66.12%, and 82.07%, respectively, with signed average forgetting of 1.83, −0.17, 1.03, and 0.52 percentage points. The results support historical retrieval retention across the evaluated backbones and four-stage sequence. The evaluation addresses sequential descriptor learning rather than online obstacle detection or explicit map-change identification.