Research On Non-cooperative Target Recognition And Relative Pose Measurement Technology For On-orbit Service

With the continuous improvement of aerospace technology,the demand for space on-orbit services such as on-orbit fuel filling,space debris cleaning,and decommissioning of obsolete satellites is increasing.Space non-cooperative target recognition and relative pose measurement are the basis of the above-mentioned on-orbit service tasks.Affected by factors such as spatial interference,illumination changes,and image quality,the accuracy and stability of traditional algorithms and deep learning algorithms in spatial non-cooperative target recognition and pose measurement still need to be improved.In response to the above problems,the paper carried out the following work:First of all,for the space non-cooperative target observation task,the noncooperative target measurement model and the construction of the point cloud recognition library are completed.Considering the characteristics of strong antiinterference performance,low cost and high recognition accuracy of TOF cameras,this paper uses the combination of open source satellite point cloud model and selfcollected data to construct a data set to form a spatial non-cooperative target point cloud recognition library,a space non-cooperative target recognition and measurement model based on TOF camera is established.Secondly,a recognition algorithm based on AL-Point Net++ is designed for the point cloud recognition task of non-cooperative objects in space.For the problem that the Point Net++ network is highly dependent and the ability to recognize complex targets is relatively poor,the Attention-LSTM network structure is integrated,and the AL-Point Net++ algorithm is proposed to optimize the extraction effect of feature points,and the efficiency and accuracy of point cloud recognition are obtained.effectively improved.Thirdly,in order to meet the requirement that the relative pose measurement accuracy of non-cooperative objects in space still needs to be improved,the accuracy of the point cloud registration algorithm is optimized to complete the relative pose measurement of the point clouds of non-cooperative objects in space.The eigenvectors and transformation relations are obtained by the principal component analysis method,the matching effect is checked by the RANSAC algorithm to ensure the accuracy of the coarse registration,and the rotation matrix and translation matrix are obtained by improving the fine registration algorithm of ICP.Experiments show that compared with the traditional algorithm,the registration effect of this algorithm is better,and more accurate relative pose measurements can be obtained.Finally,a ground-scale verification scene for space non-cooperative target recognition and relative pose measurement is built to verify the effectiveness and accuracy of the above algorithm.The TOF camera is anti-interference,and it is not affected by the surrounding lighting environment when extracting the point cloud of the target.The experimental results show that the proposed algorithm has good accuracy for the space target scale model,quarter-sat satellite recognition and relative pose measurement,and also has good effectiveness and accuracy in the dual quarter-sat satellite docking scenario.The experimental results show that the AL-Point Net++ algorithm proposed in this paper has higher recognition accuracy and real-time performance.Compared with the traditional algorithm,the accuracy is increased by 4.75%,and the training time is shortened by 7.98%.The angle error of the relative pose measurement algorithm based on point cloud registration is less than 0.4°,and the displacement error is within 3mm,which can accurately identify and accurately measure space noncooperative targets in a short time.