Landmark-based RADAR SLAM for Autonomous Driving

Simultaneous Localization and Mapping (SLAM) is a ubiquitous problem in the field of autonomous driving industry. It provides vehicles with an essential capability to explore and navigate in unknown environments autonomously. This work solves the SLAM problem using Extended Kalman Filter (EKF) and Rao-Blackwellised Particle Filter (RBPF) implementations under the assumption that camera annotations are reliable enough to ease the radar data-association problem. EKF-SLAM and FastSLAM algorithms were evaluated on the real-world nuScenes dataset in point-target and extended-target scenarios and the mean estimated localization error was as low as 0.31 m. Additionally, landmark-based map was built by filtering out dynamic targets in the environment using Random Sample Consensus (RANSAC) algorithm. Various types of consistency checks were performed to evaluate the estimation performance of these algorithms.