Deep Neural Network Method for Aircraft Pose Detection During Approach and Landing Process

Visual-aided navigation provides a feasible solution for achieving high-precision landing of fixed-wing aircraft under GNSS-denied or low-visibility conditions. However, existing runway detection and key-point extraction methods still suffer from insufficient accuracy and poor real-time performance under complex environments and limited onboard computational resources. This paper proposes a runway key corner detection method based on an improved High-Resolution Network (HRNet). A lightweight HRNet is employed to extract heatmaps of runway corner points within the region of interest (ROI) during landing. Through a parallel optimization process including threshold filtering, Gaussian smoothing, local maxima detection, and sub-pixel refinement, precise corner coordinates in the image pixel coordinate system are obtained. Furthermore, by leveraging the complementary characteristics of the Perspective-n-Point (PnP) algorithm and the vanishing-point-based pose estimation method, a segmented pose estimation framework is designed to estimate the aircraft pose. Real flight experiments demonstrate that, using image information alone, the proposed algorithm achieves pose estimation errors as low as 1.18 m, 3.37 m, and 0.067°, meeting the accuracy and real-time requirements of visual-aided navigation for fixed-wing aircraft during high-speed approach and landing.