Vehicle pose estimation in cluttered urban environments using multilayer adaptive Monte Carlo localization

In this contribution, we propose multilayer adaptive Monte Carlo localization (ML-AMCL) in combination with 3D point registration algorithms as a GPS-independent framework for precise global vehicle pose estimation in challenging urban environments. Scans from a 3D LIDAR sensor are split into a set of horizontal layers which are then used for localization with separate instances of an AMCL algorithm. A consistency check is performed for the obtained pose estimates in every time step and feasible results are fused. It is shown, that ML-AMCL is superior to existing localization approaches and is well suited as a prior for 3D point registration algorithms for the refinement of the pose estimate. Our key contributions are: i) proposal of ML-AMCL. ii) Incorporation of prior information from ML-AMCL into different pose refinement procedures, i.e. ICP variants and normal distribution transform (NDT), for precise vehicle localization. By means of experimental evaluation with a challenging real data set, performance of the reference localization system is verified. The proposed localization framework achieves a mean Euclidean measurement error of 0:2m under severe adverse environment conditions. The latter include high clutter densities in the sensor measurements and semi-static objects in the localization map.