Research On Space Target Few-shot Recognition And Pose Estimation Algorithm Based On Deep Learning

With the rapid development of space situational awareness technology,spatial information is explosively increased,and multi-source heterogeneous information provided by various spatial measurement systems brings great challenges to space situational awareness development.Meanwhile,the rapid rise of artificial intelligence technology represented by data mining and deep learning brings hopes for the instant development of space information technology.How to use image data to identify and further estimate the position and attitude of space target has become a hot topic in recent years at home and abroad.However,the main issues of this research are as follows: first,the current space target recognition and pose estimation mainly rely on the traditional image processing algorithm and geometric optimization algorithm,resulting in a semantic gap and poor generalization abilities.Second,due to many factors such as the space target complicated imaging environment,the available valid data is limited.For example,the number of samples is a few and cooperative markers of some targets cannot be known,such as optical targets,so specific information cannot be accurately known.To address above issues,this thesis proposes and implements few-shot space target recognition and pose estimation algorithm based on deep learning.Firstly,this thesis presents a few-shot recognition algorithm based on transfer learning.This algorithm is implemented by an end-to-end few-shot deep learning framework,mainly including a deep embedding module and a metric module.Because space targets actually belong to fine-grained apparel recognition with high intraclass variance and subtle interclass distinctions,the center loss and soft-max loss to jointly supervise the deep embedding module according to intraclass compactness principle,and study their balance factor,so that the model can learn discriminative deep features.In addition,for the nearest neighbor metric module,the global pooling descriptor is added into each local descriptor to reduce interference from space target background noise,thus enhancing the model robustness.Extensive experiments demonstrate that the proposed algorithm this thesis outperforms traditional space target recognition algorithm and recent state-of-the-art few-shot recognition algorithm.,and compared with latest DN4 few-shot model,proposed model gain 2.35% improvements in space target BUAA data set.Next,this thesis proposes a monocular space target pose estimation algorithm based on landmark detection.The main processes include pose information to two-dimensional image mapping,space target detection,landmark detection and Pn P geometric algorithm.Space target have multiple scales caused by different shooting distances,so we choose high resolution and multi-scale target prediction network to replace classic feature pyramid network in the target detection stage.For landmark detection model,through the combination of high-resolution network and online hard landmark mining algorithm,proposed model can focus on a part of obscured landmarks due to change of space target posture.Besides,by adding dilated convolution to increase perceptive field in the feature fusion stage and use global information to infer obscured landmark.Extensive experiments demonstrate that proposed algorithms in this thesis can effectively reduce the error rate of pose estimation,and compared to latest pose estimation ESA model,our proposed model can reduce 1.68% error rate in speed data set.The proposed few-shot recognition and pose estimation algorithm studied and implemented can be applied to space situational awareness.They can be applied to detect and track space target of interest(including space stations,satellite,launch vehicles,and its broken pieces,etc.)as well as space proximity missions(spacecraft docking and communications,debris removal).It has important theoretical significance and practical value for future space inorbit service.