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

During the approach and landing phase of fixed-wing aircraft, especially under GNSS-denied or low-visibility conditions, vision-based autonomous navigation is attracting increasing attention and is expected to play a key role in enhancing flight safety. However, existing methods often struggle to provide stable and accurate pose estimates because of the high-speed motion of fixed-wing platforms and the low-texture, low-contrast characteristics of infrared imagery. To address these challenges, this paper proposes LandNavNet, a geometry--learning hybrid pose estimation network tailored to approach and landing scenarios. LandNavNet tightly couples runway corner detection with a Singular Value Decomposition (SVD)-based rigid pose solver to compute physically consistent six-degree-of-freedom (6-DoF) poses directly from infrared images. The network incorporates a six-corner runway geometric model and a lightweight keypoint regression module that explicitly predicts runway corner locations with high precision, thereby enhancing robustness in challenging environments. An SO(3)-constrained SVD pose estimation module is further employed to derive optimal rigid transformations, preserving the geometric interpretability of analytical solvers while leveraging the scene generalization capability of deep learning. Experiments on the Approach and Runway Dataset (ARD) demonstrate that LandNavNet delivers stable and accurate pose estimates for high-speed fixed-wing approach and landing, with overall performance meeting the accuracy requirements for fixed-wing automatic landing. These results suggest that LandNavNet provides an effective technical pathway toward high-precision vision-based navigation of fixed-wing aircraft in complex operational environments.