Runway-geometry-aware infrared monocular pose estimation for fixed-wing approach and landing

When satellite navigation is unavailable, vision-based pose estimation can provide navigation guidance for fixed-wing aircraft during approach and landing. However, high-speed motion and sparse image textures make stable and accurate six-degree-of-freedom pose estimation challenging. To address these challenges, this paper proposes LandNavNet, a monocular pose estimation method that combines deep-learning-based runway landmark perception with geometric pose estimation. The detections produced by a lightweight network are structured as a six-node runway topology graph. The graph refines observations of four runway corners and two endpoint-derived midpoints while explicitly preserving the midpoint constraints. The network predicts a positive-definite anisotropic covariance for each runway landmark and uses these covariances in weighted pose optimization. Considering the approximately planar geometry of a runway, LandNavNet ranks valid planar pose hypotheses according to the visual consistency between adjacent frames and selects the final pose solution. Experiments on a real-flight approach sequence show that the East, North, and Up root-mean-square errors (RMSEs) are 1.310, 0.938, and 0.685 m, respectively. The corresponding yaw, pitch, and roll RMSEs are 0.489°, 0.937°, and 0.178°. Across three simulated scenarios, the maximum position RMSE is 3.114 m, while all attitude RMSEs remain below 1°. LandNavNet achieves a processing rate of approximately 35 frames per second on the RK3588 embedded platform. These results indicate that LandNavNet provides favorable accuracy, real-time performance, and cross-scenario adaptability under the evaluated conditions, providing a basis for further research and engineering validation of vision-only navigation during fixed-wing approach and landing.