失效卫星远距离相对位姿估计与优化方法

针对远距离非合作慢旋目标位姿估计精度问题，提出一种融合图像超分辨与视觉SLAM的相对位姿估计方法。算法主要包含3个步骤：通过梯度引导生成式对抗超分辨技术，提升目标图像的质量以获取更多更高质量的特征点；构建特征数据库实现当前帧与特征数据库的匹配，提升旋转目标的特征跟踪稳定性；利用图优化对多帧图像进行联合位姿优化，消除累计误差，得到更为精确的估计结果。为稳定网络的训练，将自然进化算法引入到对抗训练中。为增强模型的泛化性和鲁棒性，实验中的数据集采用半物理仿真获得。实验结果表明，当等效距离为25 m且失效卫星以25（°）/s的速度旋转时，目标图像经超分辨网络增强后，能够实现连续稳定的长时间测量。

Long-range relative pose estimation and optimization of a failure satellite

To improve the accuracy of pose estimation of long-range non-cooperative targets with slow rotation, a method for relative pose estimation is proposed based on fusion of image super-resolution and visual Simultaneous Localization And Mapping (SLAM). The method mainly includes three steps. First, a gradient guidance Generative Adversarial Network (GAN)-based super-resolution model is utilized to improve the quality of images, so as to obtain more and higher quality feature points. Second, a feature database is constructed to match the current frame with the feature database, so as to improve tracking stability of the rotating target. Thirdly, pose graph optimization is carried out in multiple frames to optimize the joint pose, so as to eliminate the cumulative error and obtain more accurate estimation results. To stablize the training of GAN, an evolutionary algorithm is introduced. To enhance the generalization and robustness of the model, the dataset is obtained by semi-physical simulation. Experimental results show that when the imaging distance is equivalent to 25 m and the target is rotating at 25 (°)/s, our algorithm can realize continuous stable measurement after the images are enhanced by the super-resolution model.