关键词: 失效卫星, 旋转目标, 生成式对抗网络, 低照度图像增强, 小样本, 相对位姿估计

Abstract: The serious damage of information in images under a low-light imaging condition may reduce the accuracy and robustness of the pose estimation of a failure satellite. Thus, this paper proposes an unsupervised Generative Adversarial Network (GAN)-based low-light image enhancement model. The generator is based on the U-Net with dense residual connection. The discriminator is designed as a global-local structure, and the traditional single scalar output of the discriminator is extended to the multiple scalar output. The single natural image Generative Adversarial Network (SinGAN) is improved based on evolutionary and parallel training methods to augment the few-shot training samples. Finally, a local map of feature information is established based on initialization of ORB-SLAM to alleviate dependence of pose estimation on reference frame. Through dense matching of Region of Interest (ROI) in key-frames, a nonlinear optimization model of the normal vector of plane and the mounting height of monocular camera is established to solve the scale factor. Through the similarity transformation after closed-loop detection, an equation for combined pose optimization in key-frames is established to achieve global correction of the pose transformation matrix. The experimental results show that the maximum error of the attitude angle is 4° in low-light images, while the maximum error of the attitude angle is 0.5°in enhanced images. When the target is rotating 5 cycles at 20°/s, our method can track 5 cycles in relatively static state, and 4.5 cycles in the range of 1 m horizontal maneuver. Hence, the proposed method can satisfy the need for relative pose measurement of failure satellites.

Key words: failure satellite, rotating target, generative adversarial network, low-light image enhancement, few-shot, relative pose estimation