Leveraging Probabilistic Meshes for Robust LiDAR Mapping

Although a good variety of successful LiDAR-based mapping schemes have been developed, these methods present shortcomings when mapping geometrically poor environments. In these scenarios, the chosen map structure and the consideration of map uncertainty are particularly relevant for providing a robust robot motion estimation, critically affecting the quality of the resulting map. This paper introduces the use of probabilistic 3D triangle meshes in LiDAR-based mapping. Our approach combines: i) meshes, which consistently represent planar surfaces and enable the use of decimation techniques to reduce the influence of the measurement noise in the map and improve map fidelity, while strongly reducing the map size; with ii) a probabilistic on-manifold formulation of planar objects, which naturally reflects the measurement uncertainty in the mesh map avoiding inconsistencies in state estimation. The proposed methods are experimentally evaluated both individually and jointly integrated in a generic mapping scheme in different scenarios, showing the improvement in robustness and accuracy in geometrically poor environments and providing strong reductions in map size over existing schemes. We release the used datasets and C++ implementations of the proposed methods.