Skip to content

GNSS-aided 3D LiDAR SLAM (Lego-LOAM + GPS)

Date: December 5, 2023.

Summary

The Lego-LOAM was extended with an absolute GPS unary factor in a factor-graph back end to curb long-term drift—especially in the z-axis—across real-world routes. The approach builds on ideas from Lego-LOAM, LIO-SAM, and hdl_graph_slam and shows reduced drift in evaluation sequences.

Motivation: why fuse GPS with LiDAR?

Different sensors fail differently; leveraging their complementarity improves state estimation. GPS tends not to accumulate drift but can jump due to interference/obstruction; LiDAR odometry accumulates drift but avoids abrupt jumps and prefers dense environments, struggling in open spaces.

System overview

A modern LiDAR SLAM stack separates a front end (LiDAR odometry) and a back end (graph optimization with solvers like g2o, Ceres, or GTSAM). Sensor data flows through odometry to produce a SLAM estimate, with the back end refining poses.

Work is based on Lego-LOAM and incorporate lessons from LIO-SAM and hdl_graph_slam.

GPS integration design

Factor type and coordinate frame

  • Treat each GPS measurement as a prior / absolute / unary observation (a GPSFactor).
  • Convert LLA → local Cartesian (e.g., ENU) using GeographicLib before adding to the graph.

Optimization engine

  • Using GTSAM/iSAM2 avoids hand-deriving residual Jacobians (in contrast to defining them directly in Ceres).

Practical workflow

  1. Compare measurement covariance; set the GPS factor noise.
  2. Convert GPS to the local Cartesian frame.
  3. Add factor to the graph.
  4. Run iSAM2 update after adding LiDAR odometry factor(s) and the GPS measurement.

Robustness detail

  • If the GPS “jumps”, give time for re-initialization — be distrustful of initial measurements before resuming normal fusion.

GPS modalities referenced

  • PPP and RTK are noted; RTK can reach ~10 cm accuracy but requires base-station correction.

Threads in Lego-LOAM

  • LiDAR odometry (main thread), loop closure, and visualization threads are part of the baseline system architecture we extend.

Datasets & evaluation

UTBM (EU long-term autonomous driving dataset)

  • Heterogeneous sensors and long-term sequences (~5 km, road loop) captured in downtown and suburban Montbéliard (France).
  • Our comparisons show reduced drift in the z-axis versus baseline Lego-LOAM.

Mulran (multimodal range dataset)

  • Radar + LiDAR, multiple cities/environments, month-level temporal gaps, multiple loop candidates; sequences include Riverside, 6.8 km road loop.
  • Comparative overlays vs. baseline Lego-LOAM are presented.

Credits & References

  • Project presented along with Zhexin (Jason) Xu.
  • Lego-LOAM — Shan et al., IROS 2018.
  • LIO-SAM — Shan et al., IROS 2020.
  • hdl_graph_slam — Koide et al., IJARS 2019.
  • GeographicLib — coordinate conversion utility.
  • UTBM — Yan et al., IROS 2020 (EU long-term dataset).
  • Mulran — Kim et al., ICRA 2020 (multimodal range dataset).

Back to Projects