FMCW-Based Odometry for Low-Altitude UAVs: Velocity-Driven Fusion and Sensor Evaluation

Low-altitude UAV navigation by FMCW sensors is challenged by directional measurement degradation, leading to unreliable velocity estimation and degraded odometry accuracy. To address this issue, we propose a noise-separated FMCW velocity observation model based on Singular Value Decomposition (SVD), which separates reliable and degenerate observation directions. The resulting observations are fused within an SVD-driven error-state extended Kalman filter (ESKF), where updates are performed in the SVD observation space to prevent uncertainty from degenerate directions contaminating reliable ones. Inter-Frame Temporal Consistency for abnormal velocity rejection and an IMU-gated barometric altitude update further improve robustness during takeoff. Extensive real-world flight experiments demonstrate improved measurement Gaussianity and odometry accuracy, while providing a systematic comparison of FMCW LiDAR and mmWave radar for low-altitude UAV navigation. We conclude that both sensing modalities suffer from observation degeneration. While the dense point clouds of FMCW LiDAR still enable reliable velocity estimation, mmWave radar requires SVD-based decomposition to isolate unreliable observations. Combined with its lower payload and longer sensing range, this makes mmWave radar more suitable for practical UAV deployment.