Research On Key Point Detection And Matching Algorithms For Non-cooperative Target Pose Measurement

With the gradual maturity of recoverable rocket technology,the number of satellites discarded due to failure and end of mission will increase dramatically and become non-cooperative targets in space.Therefore,the relative pose measurement technology of non-cooperative targets used in space capture will become more and more important.However,the key point detection and binocular matching based on binocular stereo vision theory are vulnerable to environmental interference in this technology.Aiming at this problem,the following studies have been carried out:Firstly,the possibility of combining the relative pose measurement of non-cooperative targets with the Deep Convolution Neural Network(DCNN)to complete the research is preliminarily analyzed and it is used to determine the research ideas.After that,the mathematical models of object recognition,binocular stereo vision subsystem and relative pose truth value calculating subsystem are established respectively,which can be used for subsequent theoretical research and experimental platform construction.Secondly,the key point detection algorithm of non-cooperative target based on DCNN is studied.A representative satellite model is regarded as a non-cooperative target experiment object,and the solar panels and their triangular brackets are used as recognition objects.According to its structure characteristics,a key point labeling software is developed.Labeling the data of recognition objects collected in the real environment with various disturbances,so as to generate data sets and train the DCNN model.Then two kinds of information output from DCNN model are analyzed by using different algorithms.Finally,the coordinates of key points and their membership relationship with the identified objects are obtained,and the key points detection is completed.Finally,the experimental comparison and analysis with the traditional feature extraction algorithm are carried out.Then,the relative pose measurement algorithm which is suitable for the above key point detection algorithm is studied emphatically.An algorithm for indirect binocular matching of key points by matching recognition objects is proposed.And the key points of detection and matching are reconstructed by binocular stereo vision.Then the relative pose measurement theory is deduced from the two cases of short distance and medium distance,and experimental comparison and analysis are made with traditional feature matching algorithm.Finally,the above algorithm is applied to the experimental platform of the DCNN-based space non-cooperative target relative pose measurement system.The self-built experimental platform consists of three subsystems: binocular stereo vision,relative pose truth calculation and in-depth learning computer.The overall flow of the algorithm,the composition,function and relationship of the subsystems are given.Based on this platform,two comprehensive experiments are carried out to analyze the feasibility of using the proposed algorithm in relative pose measurement of non-cooperative targets,and to verify the robustness of the algorithm to background,illumination and occlusion disturbances.