| The relative pose estimation of non-cooperative space targets is a key step in a series of space short-range operation tasks,such as failed satellite cleaning,space garbage capture,spacecraft on-orbit maintenance,space station rendezvous and docking,etc.This paper mainly studies the key issues in the current data-driven monocular vision pose estimation methods for non-cooperative spacecraft based on the requirements of space close operation missions with the relative depth distance of 3 m ～ 30 m with non-cooperative spacecraft,and the application potential of data-driven spacecraft pose estimation technologies in the field of aerospace measurement,control and navigation.The specific contents are as follows:Firstly,this paper studies the contour feature extraction method of the key parts of the non-cooperative space target.A space target contour feature extraction method with strong semantic image level and weak supervision is proposed,aiming at the problems of limited space target samples,time-consuming pixel level annotation and poor robustness of image level annotation to spacecraft large angle attitude rotation scene.Firstly,the key contour of the space target is annotated at the strong semantic image level,and then the contour features of the weakly supervised image level annotation are iteratively trained through the semantic segmentation objective function.The Gaussian gray histogram is used to segment the target contour,internal texture and background,and extract the target contour features.Finally,the wavelet coefficients of speckle noise in the image to highlight the contour feature information by Nakagami model.Comparative experiments show that compared with pixel level contour extraction algorithms based on weakly supervised,the proposed method improves the key indicators of contour detection accuracy by 9% ～ 12%,and also improves the robustness to illumination change and background noise when extracting the contour features of key parts of spacecraft.Secondly,carry out research on the space mapping of two-dimensional(2D)contour features of spacecraft to the three-dimensional(3D),and the training and solution of target pose,therefore,a high-precision monocular vision residual convolution network pose estimation method based on object contour space mapping constraints is proposed.The 3D voxel features of the contour and the mapping offset information are utilized to form training constraints to reduce the position and attitude offset error.The perspective projection learning loss function is used in the process of pose training to adapt to the geometric variable scene of the space target.The relative attitude of the target is obtained by Bayesian filter minimization weighted fitting according to the angle difference between the attitude labels.The Gauss-Newton iterative principle is utilized to approximate the optimal regression of the position labels to obtain the relative position of the target.The comparison experiments show that within the observation range of 3 m ～ 30 m relative depth distance,the displacement estimation errors of X,Y and Z axes using the method in this paper are less than 366.6 mm,371.2 mm and 915.7 mm respectively,and the errors of roll angle,pitch angle and yaw angle are less than4.64°,4.91° and 6.11° respectively.Within the range of relative depth distance ≤ 20 m,the position and attitude estimation errors are less than 133.6 mm and 2.32°respectively.Compared with the pose estimation method based on end-to-end data driven,the accuracy of 3D position estimation is improved by 20% ～ 38% and the accuracy of 3D attitude estimation is improved by 42% ～ 72%.Thirdly,the research on improving the efficiency of space target pose training and solution is carried out,and an optimization algorithm of space target pose training and solution efficiency based on probability mass function is proposed to solve the problem of large amount of calculation of high-precision pose estimation methods of non-cooperative space target based on data-driven.In the initial training stage,the discreteness classification of the target pose tags is reduced,and the pose tags are iteratively trained based on the probability mass function to shorten the training cycle.In the process of solving the attitude by Bayesian filtering,the nonlinear dimension of Bayesian filtering is reduced through the probability mass posterior function and the attitude quaternion multiplicative error model to reduce the solving time.In addition,the relationship between pose estimation and network loss function is reflected as approximate Bayesian inference in depth Gauss process through mathematical derivation.The influence of key network parameters on pose estimation accuracy and calculation consumption of space targets is explored,and the parameter setting range of pose estimation network suitable for space targets is proposed.Comparative experiments show that the average single pose estimation time of this method is 37% ～ 51% less than that of the end-to-end data-driven methods.Finally,a ground semi-simulation test system is built in a aerospace institute to verify the rationality and effectiveness of the proposed method in this paper for noncooperative space target pose estimation.The experimental results show that within the observation range with relative depth distance 3 m ～ 20 m,the estimation errors of each axial displacement and each attitude angle of the non-cooperative space target using the pose estimation method in this paper are less than 157.5 mm and2.86°.Within the range with depth distance of 20 m ～ 29 m,the displacement estimation errors in the X,Y and Z axes are less than 359.8 mm,362.8 mm and920.7 mm respectively,and the standard deviation is less than 15.28 mm,15.67 mm and 28.69 mm respectively.The estimation errors of roll angle,pitch angle and yaw angle are less than 4.45°,4.55° and 6.34° respectively,and the standard deviation is less than 1.171°,1.236° and 1.963° respectively.The estimation accuracy of each component of space target position and attitude meets the requirements of the space project.When Gaussian noise with mean value of 0 and variance of 0.0001 ～ 0.01 or salt and pepper noise with intensity ratio of 0.001 ～ 0.1 is added to the collected image,the accuracy of pose estimation still meets the project index requirements. |
