Research On Detection And Tracking Algorithm For Drogue Based On Vision

Air refueling technology can effectively enhance the endurance operation ability of aircraft,is a multiplier of air power,and plays a vital role in improving the combat capability of manned and unmanned aerial vehicles.At present,the aerial refueling technology of manned aerial vehicles is difficult and dangerous to operate,and the relative position of the refueling drogue can not be obtained accurately and real-time through the visual perception of pilots.Therefore,this paper studies the detection and tracking of oil drogue based on visual method.Firstly,based on histogram segmentation and cascade feature classification,the drogue target detection algorithm is studied.Through the problem of poor real-time performance of traditional target detection algorithm and combined with the shape structure characteristics of drogue,an adaptive threshold segmentation algorithm is proposed.Cascade classifier is established based on the shape characteristics of the target to extract candidate regions of drogue.The edge feature description model of drogue is established by using HOG features and SVM classifiers,and the final one is obtained by modifying the local detection region algorithm.Target detection area ensures high real-time and accuracy of drogue detection algorithm.Secondly,this paper studies the target detection method of oil drogue based on deep learning image processing.By analyzing the shortcomings of traditional methods in robustness and combining with real-time YOLO target detection algorithm,a Dapper-YOLO target detection algorithm is proposed to improve the accuracy of target detection.For the case of small targets,this paper proposes a method of combining Dappers-YOLO target detection algorithm based on cascade structure with block detection algorithm,which improves the detection accuracy and success rate in the case of small targets.Thirdly,the scale-adaptive drogue target tracking algorithm is studied.By analyzing the problems of scale fixing in correlation tracking and the inability to judge the tracking failure,this paper proposes a scale adaptive correlation tracking algorithm framework based on feature point detection.Two methods of feature point detection based on polar coordinate transformation and feature point detection based on convolutional neural network are proposed respectively from the two directions of traditional image processing and deep learning,and combined with correlation tracking.The algorithm and the Dapper-YOLO target detection network realize the drogue target tracking algorithm with high accuracy,high robustness and high real-time performance.Finally,based on the collected aerial refueling video data,this paper establishes a drogue target detection and tracking test data set for various scenarios,and labels the real reference value of the data set,and tests and evaluates the performance of the proposed method and the existing related methods from different angles.The experimental results verify the validity of the research methods and play an important supporting role in the research work of this paper.