Research On Autonomous Aerial Refueling Recognition And Measurement Technology Based On Convolutional Neural Network

Aerial refueling,as one of the key ways to effectively increase the combat radius of aircrafts,has undergone the development of manual training docking and surveillance-assisted docking.Autonomous aerial refueling(AAR)technology,as the development direction of aerial refueling and the key technology to support future unmanned aerial vehicles to complete aerial refueling tasks,is still in its infancy in China.In complex air environment,the measurement of relative pose between aircrafts is one of the key technologies to accomplish AAR tasks.Characteristics of high danger and high environmental complexity of AAR tasks place higher demands on the robustness,range of measurement,accuracy of measurement and the power of computing in measurement system.In this research,guided by the realization of relative pose measurement for AAR,after in-depth research and analyzing of the existing measurement methods,a recognition and measurement system for AAR tasks has been implemented based on convolutional neural networks(CNNs),with innovation and improvement on some key technologies.The main work of study is as follows:Firstly,a mathematical model of relative pose measurement system is constructed for AAR based on monocular vision and without cooperation marks in this subject.And the operation framework is established for the measurement system.A systematic calibration of the visual measurement system is performed to obtain the system parameters with referring to Zhang's calibration method,which lays the foundation for the relative pose estimation.At the same time,drogue recognition and tracking datasets are constructed for AAR based on the idea of supervised learning.A data augmentation method was proposed to provide effective training datasets for recognition and tracking of drogue based on the relative motion between the drogue of refueling and fuel receiving aircraft,and the environment conditions.Secondly,the drogue recognition and localization methods based on arc-level features and multi-receptive field CNN were proposed according to the key technical requirements of initial drogue acquisition phase of the measurement system.The method of recognition and localization of drogue based on arc-level features uses arc features,multi-features fusion,and random sampling consistent contour fitting to achieve robust recognition and precise positioning of the drogue.Compared with existing methods,the robustness and real-time performance are significantly improved.To improve environmental adaptability and anti-interference ability,multi-receptive field CNN and multi-scale residual prediction are proposed,by which the average positioning accuracy reaches pixel-level and recognition rate reaches 95%.In the stage of stable tracking and local detection of drogue,this paper converts the tracking problem to a matching problem based on the Siamese network algorithm.The robustness of drogue tracking is guaranteed by improving basic feature extraction network and the proposed online layer-wise feature modulation.The tracking accuracy is improved by multi-scale reverse and stepwise fine-tuning of the bounding box.In the stage of local detection,long-term stable tracking of drogue is achieved by the proposed search area selection strategy and the recognition and localization of multi-receptive field network method.In addition,in order to meet the requirements of robust high-precision relative pose estimation in large-distance range in the phase of relative pose estimation,a novel relative pose estimation method based on back projection features is proposed.Robust high-precision relative pose estimation is achieved by extracting accurate structural features,reversed solution of 5 DOF spatial pose,and state estimation based on the motion model.Compared with the existing algorithms,the measurement range,accuracy,and robustness are significantly improved.As the key part of the entire measurement system,the proposed method guarantees the accuracy and robustness of the measurement system.Finally,according to the research of the above key technologies,various sets of experiments such as dataset testing,numerical simulation experiment,and ground experiments for each key part were carried out.These experiments are carried out to analyze the performance of the proposed key technologies in the subject.The robustness,accuracy,and real-time performance of each part of the method were evaluated from various aspects such as complex lighting,movement state,and weather conditions.The robotic arm joint experiment of the entire system was performed to analyze the feasibility and practicability of the proposed measurement system.A complete relative pose measurement technical route is proposed for AAR technology of China.