Pose Estimation Of Non-Cooperative Spacecraft Based On Neural Network

Pose estimation of spacecraft has a wide range of applications in space missions such as manned aerospace,space station rendezvous and docking,space debris removal,spacecraft in-orbit service,and celestial landing.Among various spacecraft pose estimation methods,the pose estimation method of non-cooperative spacecraft has attracted much attention because of its robustness to occlusion and illumination and wider application range.This paper proposes two pose estimation methods based on neural network,which have high estimation accuracy and good robustness to illumination and occlusion for non-cooperative spacecraft pose estimation.The first pose estimation method proposed in this paper is an end-to-end pose estimation based on neural networks.This method uses the image as the input of the neural network after preprocessing,and then the neural network outputs the initial value of the predicted pose corresponding to the image.According to the initial pose value,the algorithm of minimizing the geometric residual of the projected contour of the spacecraft CAD model in the image is used to optimize the pose parameters of the spacecraft in the image,and finally the maximum likelihood estimation of the spacecraft pose in the image is obtained.In the simulation experiment of the deep space background datasets,the average relative position error of the pose parameters estimated by this method is 4.06%,the average Euler angle error is 10.85°,and the average pass rate reaches 73.16%.The experimental results show that the method has the characteristics of high accuracy and good robustness to illumination and self-occlusion in the deep space background condition.However,when the images have complex background noise,the estimated pose parameters of this method have large errors,and the average pass rate is only 32.44%.Therefore,the second pose estimation method is proposed in this paper to solve the above problem.The second pose estimation method proposed in this paper is pose estimation based on key-points.This method first selects a series of spatial points on the spacecraft shape as key-points,and generates a descriptor for each key-point.Then,the projection point coordinates of the spatial points in the image are obtained according to different pose labels,and the projection point coordinates and the corresponding descriptors are used as the data labels in the offline training step of the method to train the neural network.In the running step of the method,the spacecraft image is input into the neural network,and the neural network outputs the descriptor of the feature point and the position of the feature point in the image.Next,we use the descriptor to establish the matching relationship between spatial points and image feature points,and use the combination of RANSAC and EPn P to solve the spacecraft attitude according to the matching relationship.Extensive experiments show that the pass rate of this method for pose estimation of deep space background images and earth texture images is 71.04% and 77.64% respectively.It is worth noting that in the deep space background images,the pass rate obtained by this method is not much different from that of the first method,but in the earth texture images,the pass rate obtained by this method is greatly improved.The above experimental results show that the proposed method can not only have high pose estimation accuracy for spacecraft images with background noise,but also accurately estimate the spacecraft pose in the deep space background.The innovations of this article include:（1）Generate a large number of datasets for simulation experiments to train and evaluate the performance of pose estimation algorithms;（2）Design the label normalization method to solve the gradient imbalance problem in end-to-end pose estimation,and use a new image preprocessing method that can highlight the outline of the spacecraft and improve the accuracy of pose estimation;（3）Two neural network frameworks are proposed,namely,and end-to-end pose estimation framework and key-points detection and matching framework;（4）An optimization method is proposed to minimize the geometric residual of the projected contour of the spacecraft model,which can further improve the accuracy of attitude estimation in the deep space background.