TE-ISC-PGO: A Tunnel-Enhanced LiDAR–Inertial SLAM Framework for Underground Mine Roadway Mapping

Abstract

LiDAR–inertial simultaneous localization and mapping (SLAM) is fundamental to autonomous robots operating in underground mine roadways, where corridor-like geometry, repeated structures, and weak observability can lead to accumulated drift and perceptual aliasing. This article presents TE-ISC-PGO, a tunnel-enhanced backend built on a FAST-LIO2 front end. The backend combines anisotropy-aware intensity-context matching, degeneracy-aware keyframe selection, adaptive loop-constraint weighting, and confidence-weighted roadway structural regularization to improve loop-constraint quality, reduce graph redundancy, and stabilize weakly observable lateral and yaw directions. The system was evaluated on a 654.48-s underground round-trip sequence using an independently generated external-map NDT–UKF trajectory as the reference. Under the algorithm-native output protocol, TE-ISC-PGO achieved an APE RMSE of 1.9308 m, compared with 2.0477 m for FAST-LIO2, 2.3927 m for SC-PGO, and 2.3341 m for ISC-PGO. Because the compared methods produce trajectories at different output densities, a secondary evaluation was conducted at 6,519 common timestamps. Under this density-controlled protocol, TE-ISC-PGO and FAST-LIO2 achieved comparable APE RMSE values of 2.0618 and 2.0353 m, respectively, while TE-ISC-PGO obtained a slightly lower translational RPE@1 m of 1.6707 m versus 1.6887 m. Independent ablations further showed that the full configuration reduced the keyframe count from 522 to 286 while improving APE RMSE, APE median, and revisit RMSE relative to the Base backend. Overall, the results demonstrate a tunnel-specific, graph-efficient backend that improves native-output global trajectory agreement, substantially reduces graph redundancy, and maintains competitive density-controlled trajectory consistency on the evaluated underground route.