Abstract: To address the degradation of single-frame accuracy and the instability of dynamic pose estimation caused by visual occlusions during autonomous aerial refueling docking, including probe occlusion and partial out-of-view drogue observations, an occlusion-robust monocular dynamic pose estimation method for refueling drogues is proposed. The method adopts a cascaded framework of monocular perception, single-frame pose solving, and multi-frame dynamic correction. YOLOv8-seg is first used to extract the drogue region and edge masks. Then, robust ellipse fitting is performed to obtain the initial pose, and nonlinear optimization is employed to minimize the reprojection error, thereby generating high-accuracy single-frame pose observations. Furthermore, an adaptive unscented Kalman filter is introduced for robust multi-frame pose estimation. By constructing a reduced-order 10-dimensional state-space model and a measurement-residual-driven adaptive covariance adjustment mechanism, the continuity and stability of dynamic pose estimation under visual occlusion conditions are improved. Experimental results show that the proposed method achieves a stable processing speed of more than 45 FPS on the RK3588 embedded platform, and the relative pose measurement error under visual occlusion conditions is less than 8%. Compared with the single-frame geometric pose-solving method and the fixed-noise unscented Kalman filtering method, the proposed method exhibits better temporal stability and disturbance rejection capability, satisfying the real-time pose measurement requirements in the docking stage of autonomous aerial refueling.

Key words: erial refueling, monocular vision, drogue pose estimation, visual occlusion, adaptive unscented Kalman filter, YOLOv8-seg