SLAM for Infrastructure
about.
This peer-reviewed study surveys and benchmarks vision-, LiDAR-, and LiDAR-visual-inertial SLAM specifically for enclosed and GPS-denied (EGD) infrastructure environments (e.g., bridges, tunnels, underground utilities). It categorizes methods chronologically, explains core SLAM components (front-end/back-end, loop closure, optimization, fusion), and evaluates 11 open-source ROS implementations (VINS-Mono, ORB-SLAM2, LIO-SAM, Fast-LIO2, SC-Fast-LIO2, LeGO-LOAM, SC-LeGO-LOAM, A-LOAM, LINS, F-LOAM, LVI-SAM) using our dataset and a public EGD dataset. The review also covers NeRF-based SLAM trends and practical metrics for field use.
challenge.
Infrastructure inspections often occur in EGD settings with poor lighting, sparse features, and no GPS, where drift, tracking loss, and point-cloud distortion can undermine autonomy. Prior surveys rarely center EGD conditions or provide head-to-head benchmarking, and practitioners face trade-offs: vision is lighter and cheaper but brittle in low light; LiDAR is robust but power- and compute-hungry; fusion choices (tight vs. loose) are nontrivial. A consolidated, inspection-focused map of the method space and its practical failure modes and metrics was missing.
results.
LiDAR-based SLAM excels in EGD environments; however, some vision methods remain viable under EGD with proper conditions. LiDAR-visual (and visual-inertial) fusion typically outperforms either modality alone.
The paper details scan matching (ICP, GICP, NDT, KISS-ICP, CT-ICP, X-ICP, GenZ-ICP); feature pipelines (SIFT/SURF/ORB/FAST/Harris); and optimization/fusion stacks (EKF/UKF, pose graphs, GTSAM/g2o/Ceres). These choices materially impact drift and robustness in EGD.
A benchmark of 11 ROS packages on our dataset and a public EGD set compares accuracy and compute using trajectory/drift RMSE and point-cloud distortion considerations—giving practitioners realistic expectations for enclosed sites.
Actionable guidance: pick LiDAR or fused pipelines for high-stakes EGD runs; consider vision(VINS/ORB-family) where lighting and texture can be controlled; use loop-closure and tight fusion to contain drift over long traverses.
testimonial.
“This review connects SLAM theory to field reality for GPS-denied inspections. It compares the systems we actually deploy—vision, LiDAR, and fused—under enclosed conditions, and translates the results into clear guidance for choosing and tuning a stack.”
Ali Ghadimzadeh
Ph.D. Candidate, Mechanical Engineering
