Monocular-based Pose Estimation For Non-cooperative Space Targets Based On Deep Learing

Due to the advantages of non-contact,small sensor size,low power consumption,and rich of information,along with the steady development of deep learning technology in the field of computer vision,and the wide application in practical engineering tasks such as object recognition,scene segmentation,scene classification,etc.Non-cooperative spacecraft monocular visual relative position and attitude measurement method based on the deep learning technology has become a mainstream due to its wide range of application scenarios in spacecraft rendezvous and docking,spacecraft on-orbit maintenance,and space debris removal.One of the important options in the task.This paper takes non-cooperative spacecraft as the research object,adopts monocular visual measurement,takes pictures taken by RGB cameras as input,uses neural network to assist with Pn P algorithm,and realizes the solution of the relative pose between non-cooperative spacecraft.For non-cooperative spacecraft images,the target scale changes greatly,and the image background(clouds,etc.)is complex.A dual-channel neural network is designed with reference to VGG and Dense Net networks.First,using the VGG network structure as a blueprint,the network is artificially disassembled into two channels,which are used to characterize the characteristics of two granularities in the image.Then,referring to the design idea of the Dense Net network,the output between the upper and lower channels,the output of different convolutional neural network modules in the same channel,and the tensor corresponding to the original image are combined to form a composite tensor,which is used as the input of the residual neural network.,To identify the corresponding pixels in the image of the positioning feature points on the target non-cooperative spacecraft,and give the corresponding pixel coordinates.Aiming at the calculation real-time requirement in the relative position and attitude measurement of non-cooperative spacecraft,considering the problems of many neural network parameters and large amount of calculation,a pruning method of neural network is proposed to realize the lightweight of the network.By calculating the partial differential of the elements in the feature tensor output by the convolutional neural network,and taking the arithmetic mean of the absolute values of the partial differentials of the elements in the same channel,the network is artificially deleted in the order of the arithmetic mean from small to large.In order to improve the accuracy of the relative pose of the overall solution.Using the Pn P algorithm,the pixel coordinates of the feature points on the target noncooperative spacecraft output by the neural network are used as input,and the relative pose of the target non-cooperative spacecraft body coordinate system relative to the observation satellite satellite camera coordinate system is The output solution method.Using the calculated relative pose and the actual relative pose to construct a new objective function,and under the premise of keeping the neural network structure unchanged,the neural network is trained in optimization to improve the accuracy of the relative pose of the overall output of the scheme.