DRGC-LIO: a dynamic removal and ground constraint LiDAR–Inertial odometry for UGVs

Abstract

LiDAR–Inertial Odometry (LIO) has been widely adopted for autonomous navigation due to its high-precision state estimation capability. However, existing LIO systems still face two major challenges. First, most methods rely on the assumption of static environments and therefore suffer from performance degradation when dynamic objects interfere with point cloud registration. Second, current ground constraint strategies usually assume globally consistent ground elevation, which limits their applicability in environments with uneven or sloped terrain. To address these challenges, this paper proposes a dynamic removal and ground constraint LIO (DRGC-LIO) framework. Specifically, a dynamic object removal method based on an improved vertical voxel occupancy descriptor is designed to efficiently detect dynamic regions by representing the differences between the current keyframe and the local submap using binary height occupancy codes. Furthermore, to mitigate the misclassification of ground points during dynamic removal, a ground recovery mechanism combining height-layer search and geometric constraints is introduced. In addition, a ground constraint model based on local consistency is developed to accommodate terrain variations and suppress vertical drift by incorporating the constraint into a factor graph optimization framework. Additional ablation, keyframe-versus-frame filtering, axis-wise error, and reproduced baseline experiments are included to quantify the contribution of the DRGC modules. Extensive experiments on public datasets (KITTI, ECMD, and UrbanNav) and real-world robotic platforms demonstrate that the proposed framework effectively removes dynamic disturbances and improves localization accuracy.