HollowNet: Real-time drogue detection and pose estimation via embedded vision for robust autonomous aerial refueling

Nowadays, unmanned aerial vehicles have a significant impact on the low-altitude economy and national defense. Autonomous aerial refueling (AAR) is a critical technology to extend drone endurance and enhance combat capabilities. However, existing methods mainly rely on two-stage frameworks. These suffer from computational redundancy and limited robustness under dynamic aerial conditions. In this paper, we propose HollowNet, a real-time embedded vision network for end-to-end drogue detection and pose estimation in AAR. The network integrates a single-stage detector with a decoupled head to simultaneously predict bounding boxes and LED keypoints on the drogue. We further optimize the object keypoint similarity loss by incorporating geometric constraints of the drogue target. A unified drogue positioning system, combining detection and pose estimation, is introduced to output precise docking coordinates. The feasibility of our system is validated on an aerial refueling virtual simulation platform, and on an RK3588-based embedded system it achieves 50 FPS, enabling real-time performance. It attains an mAP50–95 of 82.11% on real AAR datasets, delivering the best speed-accuracy balance among all benchmarks.