Probabilistic Uncertainty Modeling and Sensor Fusion in Radar SLAM
Yang Xu · 2025
Robust autonomous navigation in challenging environments requires precise and reliable Simultaneous Localization and Mapping (SLAM) solutions. Conventional visual or LiDAR-based SLAM systems struggle under conditions such as fog, dust, and low lighting, while millimeter-wave radar offers superior robustness. However, radar-based SLAM encounters challenges such as sparse measurements, anisotropic uncertainty, and multipath effects. This thesis presents a novel radar-inertial SLAM system addressing these issues through probabilistic uncertainty modeling and sensor fusion. It introduces a polar-coordinate-based un-certainty framework, more accurately modeling radar measurement characteristics. A probability-guided data association method further improves radar point matching robustness. An uncertainty-aware factor graph optimization framework dynamically integrates these uncertainties, enhancing the accuracy of state estimation. The proposed radar-inertial odometry system, utilizing Doppler velocity and high-frequency inertial measurements, demonstrating enhanced performance in most scenarios. Comprehensive experiments conducted on both self-collected and public datasets show significant improvements over existing methods. Practical guidelines, along with open-source implementations and datasets, are provided to encourage further research in robust radar-based perception for autonomous robotics.