Localization for Autonomous Driving is mainly researched on camera and LiDAR data, rarely on radar data. Applying a point cloud pose graph SLAM on a modern 77GHz automotive radar sensor-set, we realize direct radar sub-map scan matching for consistent mapping and ego vehicle localization at low speeds in parking scenarios. A spatiotemporal activation maps pre-filters clutter and noise. Matching locally overlapping radar sub-maps in open-loop sections results in a locally stiffer but drift corrected pose graph architecture. Integrating point cloud segmentation of 3d radar detections yields semantically consistent scan matching and increased static map accuracy. Our results reveal how semantic information, temporal filtering, and Normal Distribution Matching can outperform the accuracy of direct radar mapping and localization. Our approach SERALOC yields a new benchmark of 0.4 m absolute pose error (APE) over a distance of 624 m, measured on real-world test data against differential GPS data.
SERALOC: SLAM on semantically annotated radar point-clouds
2021-09-19
2262095 byte
Conference paper
Electronic Resource
English
Learning Semantics on Radar Point-Clouds
IEEE | 2021
|Landmark based radar SLAM using graph optimization
IEEE | 2016
|Simulation of automotive radar point clouds in standardized frameworks
GWLB - Gottfried Wilhelm Leibniz Bibliothek | 2021
|