kRadar++: Coarse-to-Fine FMCW Scanning Radar Localisation

Area of Science:

• Robotics and Autonomous Systems
• Sensor Fusion
• Simultaneous Localization and Mapping (SLAM)

Background:

• Current autonomous systems often rely on visual data, limiting performance in adverse weather or low-light conditions.
• Radar-based localization offers a complementary sensing modality, robust to environmental changes.
• Existing radar place recognition (RPR) and scan matching systems show promise but require integration for robust navigation.

Purpose of the Study:

• To develop and evaluate a novel two-stage system integrating radar place recognition with spectral landmark-based pose estimation.
• To demonstrate the complementary nature of RPR and scan matching for robust autonomous navigation.
• To validate the system's long-term durability and performance on an extensive urban autonomy dataset.

Main Methods:

• A two-stage system combining topological localization candidates from RPR with precise pose estimation.
• Utilizing spectral landmark-based techniques for accurate localization.
• Offline experiments on a large-scale, radar-focused urban autonomy dataset.

Main Results:

• The proposed system achieves performance comparable to state-of-the-art radar localization methods.
• Demonstrated long-term durability and reliability of the radar sensing technology for autonomous navigation.
• Achieved high recall and precision with reduced pose estimation variance over a month of trials.

Conclusions:

• The integrated system offers a robust solution for radar-based localization and navigation.
• The approach is a crucial step towards developing radar teach-and-repeat (RTR) systems.
• Enables autonomous navigation in challenging conditions with extreme appearance changes or inclement weather.