Answers are grounded in this paper and cite their sources.
Air–ground localization becomes brittle when visual–inertial drift or an abrupt state jump places a camera pose outside the attraction basin of local camera–LiDAR registration. We address this recovery problem with a coarse-to-fine cross-modal pose-realignment framework. Visual features are back-projected and evaluated against local LiDAR planes through point-to-plane residuals, avoiding learned appearance correspondences at run time. An association-quality check determines whether the propagated pose is sufficiently supported by local geometry. When the support is insufficient, a bounded particle swarm optimization (PSO) search selects a coarse pose hypothesis, which is subsequently re-fined by a visual–inertial–LiDAR factor graph. The proposed framework therefore couples implicit geometric association, bounded basin exploration, and local multimodal refinement to recover from large pose errors. Experiments use the Eu-RoC sequences with a static Leica scan and real indoor and outdoor air–ground platforms with UGV-provided maps. In the reported severe cases, the method reduces translational RMSE by 33.8% under a 0.5 m initial offset and reduces post-disturbance mean absolute position error by 55.2% after a 30◦ mid-sequence yaw jump, relative to the best successful baseline in each test. Indoor tests with one UGV and three UAVs further demonstrate 20 Hz operation in a shared LiDAR map frame.
No takes yet. Share an insight, caveat, or question.
Jian et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: