An Experiment of Cone-Sleeve Recognition in UAV Aerial Refueling using Multi-Modal ViT Technology

This study presents a Semantically-Decoupled Vision Transformer (SD-ViT) framework for robust aerial refueling drogue acquisition. To resolve the conflict between high-latency Transformers and real-time constraints, we introduce an "Offline-Prompt, Online-Vision" architecture integrated with hardware-aware tensor optimization. Flight tests under severe backlighting show that SD-ViT outperforms traditional color segmentation and provides stronger robustness under domain shift than a YOLOv8 baseline. With inference optimization, the proposed method mitigates computational bottlenecks, offering improved feasibility for onboard applications.