Physics-Informed Target Discovery for Robust Online mmWave Radar Extrinsic Calibration

Automatic extrinsic calibration of millimeter-wave (mmWave) radar is an under-explored field, constrained by the sensor’s inherent point cloud sparsity and measurement uncertainty. Existing methods often rely on specific calibration targets, require manual intervention, or fail to fully exploit the sensor’s unique physical information, leading to limitations in automation and robustness. To address this problem, this paper proposes an online, target-less extrinsic calibration method for multi-radar systems. The framework performs physics-informed target discovery by fusing Doppler velocity statistics and Radar Cross-Section (RCS) to automatically define a Region of Interest (ROI). Leveraging this dynamically defined region, the system then employs an uncertainty-aware registration algorithm (G-ICP) and an adaptive update strategy to precisely measure the sensor’s pose. Experimental results demonstrate that the proposed method achieves a Dynamic Trajectory Alignment Error (DTAE) of 1.2 cm in static scenarios and maintains a DTAE of 1.9 cm even in complex scenarios with dynamic interference. The end-to-end effectiveness of the framework is also validated in an uncontrolled outdoor environment. This work provides a practical pathway for condition-independent self-calibration of mmWave radar, highlighting the potential of leveraging intrinsic sensor physics to solve complex calibration challenges.