{RaI}-{SLAM}: Radar-Inertial {SLAM} for Autonomous Vehicles

Simultaneous localization and mapping are essential components for the operation of autonomous vehicles in unknown environments. While localization focuses on estimating the vehicle’s pose, mapping captures the surrounding environment to enhance future localization and decision-making. Localization is commonly achieved using external {GNSS} systems combined with inertial measurement units, {LiDARs}, and/or cameras. Automotive radars offer an attractive onboard sensing alternative due to their robustness to adverse weather and low lighting conditions, compactness, affordability, and widespread integration into consumer vehicles. However, they output comparably sparse and noisy point clouds that are challenging for pose estimation, easily leading to noisy trajectory estimates. We propose a modular approach that performs radar-inertial {SLAM} by fully leveraging the characteristics of automotive consumer-vehicle radar sensors. Our system achieves smooth and accurate onboard simultaneous localization and mapping by combining automotive radars with an {IMU} and exploiting the additional velocity and radar cross-section information provided by radar sensors, without relying on {GNSS} data. Specifically, radar scan-matching and {IMU} measurements are first incorporated into a local pose graph for odometry estimation. We then correct the accumulated drift through a global pose graph backend that optimizes detected loop closures. Contrary to existing radar {SLAM} methods, our graph-based approach is divided into distinct submodules and all components are designed specifically to exploit the characteristics of automotive radar sensors for scan matching and loop closure detection, leading to enhanced system performance. Our method achieves state-of-the-art accuracy on public autonomous driving data.