| In response to the needs of space target approaching navigation,grasping,and maintenance services,the 6D object pose estimation of space target is an important task for spatial autonomous on-orbit services.The visual measurement method has gradually become the most commonly used sensor for spatial target pose estimation due to its advantages such as large amount of information,low load requirement.Taking into account the characteristics of spatial target sequence images,this paper combines intelligent perception technology,deep learning technology and actual space on-orbit tasks,and proposes a network structure based on the LSTM framework,which uses inter-frame target optical flow for pose estimation.This method utilizes the time-domain information of the sequence images,and can maintain robust and accurate tracking performance even in long-term sequences.First,a method for image segmentation of spatial objects based on Fully Convolutional Networks(FCN)is studied.The FCN network is used to segment the image to achieve the acquisition of the target mask and determine the target area of interest in the image.In view of the large range of spatial image depth and scale changes,in order to reduce the loss of image detail information,the network structure is improved through the idea of hole convolution and feature layering to improve the accuracy of semantic segmentation.Then,a method for estimating the relative pose of the space target based on the dense optical flow of the target is proposed.After calculating the dense optical flow between frames of the sequence image,combine the target mask obtained by the FCN network to obtain the target dense optical flow;in order to increase the structural information of the target,the initial two-dimensional point position of the optical flow and the corresponding three-dimensional point coordinates are added to form the target frame Inter-information matrix.Using this matrix as the input,the calculation of the pose between target frames is achieved through a network structure based on MLP(Muti-Layer Perception).The optical flow extraction area is constrained by independent image segmentation to reduce the cumulative error in the traditional optical flow attitude estimation method and improve the accuracy of the attitude estimation.Finally,a high-precision spatial target pose calculation method based on LSTM is proposed.Use the long-term storage unit in the LSTM network to learn the motion regular pattern of the spatial target,avoiding the error of pose estimation caused by the missing of the target in the spatial image,local overexposure or too dark;effectively use the time domain information of the sequence image;Reduce the cumulative error in the optical flow pose estimation process;obtain more stable and accurate pose calculation results.In this paper,the synthetic data set and the standard space target data set are used for simulation experiments and verification.The results show that the method has strong robustness and high accuracy in the task of long-time sequence space target attitude estimation. |
