PolarBEVPlace: Rotation-Robust LiDAR Place Recognition via Radial Distortion-Aware Polar Transform

Abstract

LiDAR-based place recognition is fundamental to loop closure and global localization in SLAM. Bird’s Eye View (BEV) projection has emerged as an effective paradigm for 3D point cloud processing, yet rotation robustness remains expensive: prior approaches either discretize SO(2) at <inline-formula><tex-math notation="LaTeX">$\mathcal {O}(N_{R})$</tex-math></inline-formula> inference cost or require high-parameter dual-branch cross-attention. Polar BEV representations convert heading-induced 2D rotations into 1D cyclic translations, but introduce a systematic <italic>radial distortion</italic> unaddressed by existing architectures: expanding annuli with decreasing LiDAR point density cause feature activations to collapse at large radii, while global attention mixes geometrically incompatible cross-radius representations. We propose two lightweight mechanisms to address these failure modes: <bold>Radius-Aware CoordConv</bold> concatenates a deterministic radial coordinate channel to recover feature energy in sparse outer rings; <bold>Ringwise Circular Attention (RCA)</bold> confines self-attention to 1D azimuthal rings at each fixed radius, preventing cross-radius feature mixing by construction and reducing attention complexity from <inline-formula><tex-math notation="LaTeX">$\mathcal {O}((H_\theta W_{r})^{2})$</tex-math></inline-formula> to <inline-formula><tex-math notation="LaTeX">$\mathcal {O}(W_{r} H_\theta ^{2})$</tex-math></inline-formula>. The resulting single-stream network, PolarBEVPlace, achieves a mean Recall@1 of 94.0% across 14 sequences on KITTI, NCLT, and UrbanNav-HK at 2.30 ms per scan with 1.06 M parameters—a 5.8× latency reduction and 10.7× MAC reduction over BEVPlace++, outperforming or matching it on 13 of 14 sequences.