Abstract:

To solve the problem of positioning the refueling drogue during autonomous docking process in probe-drogue aerial refueling, a fast positioning scheme combining deep-learning (an improved version of YOLOv4_Tiny) and binocular vision is proposed, By inserting spatial pyramid pooling (SPP) module and modifying certain convolutional layers, the improved YOLOv4-tiny runs at 182 Hz on 416×416 inputs. The improved net is 20.47% smaller in size and 5% higher in average IoU on test set compared with the original net. Experiments of positioning are carried out with scaled model of refueling drogue. Average error of depth prediction is less than 5% and results of spatial prediction are in line with expectations. A rapid edge fitting scheme based on Yolo prediction is introduced to obtain elliptic feature of refueling drogue. Meanwhile, an augmented MRAC controller based on projection operator is established driving the receiver aircraft to track the refueling drogue. Simulation results show that the receiver aircraft tracks the drogue with an average error smaller than the capturing radius thus docking requirement is fulfilled.

Key words: aerial-refueling, deep learning, binocular vision, object detection, camera calibration, adaptive control