Real-time Detection of Aerial Refueling Drogue Based on Convolutional Neural Network

A real-time detection method of aerial refueling drogue based on convolutional neural network is proposed to address the problem of insufficient accuracy and robustness of vision-based aerial refueling drogue detection in the autonomous aerial refueling docking by drogue-probe method. Firstly, a lightweight backbone feature extraction network is designed based on the Stem module and ELAN module. Secondly, the feature focus unit is designed to enhance the deep feature space information and improve the localization accuracy of the model. Then, the attention concatenation module is proposed to correct the performance impairment of large objects caused by FPN. Finally, an anchor-free detection head is used to decompose the classification and regression tasks in the detection task and output the classification information in advance to avoid the spatial misalignment problem. Experiments demonstrate that the proposed method is more accurate than the current mainstream object detectors on the publicly available real air drogue dataset, with frame rates up to 117 FPS, which proves the effectiveness, robustness and real-time performance of the proposed method.