Research On Pose Estimation Method Of Space Target Based On Deep Learning

Spacecraft pose estimation is one of the fundamental and critical technologies for space on-orbit services,widely used in tasks such as satellite navigation and space maintenance.Spaceborne monocular cameras can capture target images in real-time with high resolution.The development of computer vision has made it possible to obtain pose information from monocular images.The spatial target pose estimation has the characteristics of large scene range,few texture features,limited image resolution and complex background environment.Conventional pose estimation methods face space targets with difficult feature extraction,depend on a priori knowledge,are computationally intensive and poorly adaptable,and thus more methods are needed to improve accuracy and enhance robustness for spatial target pose estimation.In this paper,we use the powerful feature extraction capability of deep learning to design an end-to-end method for pose estimation of spatial target image.The main research content is as follows:(1)Due to the requirement of strong spatial information extraction ability in pose estimation,the Polarized Self-Attention attention module is embedded into the residual network Res Net-50 to enhance the network’s ability to learn spatial features.The FRe LU activation function is introduced into the residual block to activate spatial-insensitive information in the network without changing the convolution.As the location and orientation information of spatial targets are independent of each other,so decoupled into two different network branches to obtain their respective information through regression.Specifically,the location is regressed through fully connected layers,while the orientation information is regressed through soft-assignment encoding.(2)To address the characteristics of pose estimation for spatial targets,including small target pixel occupancy,occlusion in the image,and sensitivity of posture rotation prediction to image resolution,this paper proposes to combine the high-resolution network HRNet with the Hybrid Dilated Convolution(HDC)module to reduce the loss of image resolution caused by downsampling.The fusion and expansion of the different branches in the HRNet output stage using the HDC module for the HRNet output stage enlarges the receptive field and improves the accuracy of pose estimation,which performs well in both simple and complex background images.Finally,the URSO space target dataset is used to verify the algorithm,and the visual prediction results and analysis of location and orientation are given.The new algorithm can effectively improve the pose estimation accuracy,is robust to spatial target pose estimation.