一种临近空间飞行器模糊星图端对端识别方法  

A Smearing Star Map End-to-end Recognition Method for Near-space Flight Vehicles

作　　者：张源 李睿 许志[1] ZHANG Yuan;LI Rui;XU Zhi(School of Astronautics,Northwestern Polytechnical University,Xi’an 710072,China)

机构地区：[1]西北工业大学航天学院,西安710072

出　　处：《宇航学报》2025年第1期92-107,共16页Journal of Astronautics

摘　　要：针对临近空间高速、高动态飞行器在大角速度机动下,星敏感器拍摄的星图严重退化致使星图识别成功率急剧下降问题,提出了一种基于深度学习的“识别—粗定位—复原—精定位”一体化星点提取方案。针对传统星点提取算法在载体高动态运动环境下,星点质心提取误差大的问题,设计了一种基于甚快区域卷积神经网络(Faster R-CNN)的主星模式分割算法。首先得到星点的粗略位置,然后通过稠密连接结构的局部星图端对端复原网络,得到星点精确质心位置。考虑真实星图信噪比极低且噪声特性复杂,采用基准星辅助的容错检测算法对星点精定位结果予以校验,剔除误识别及定位误差过大的星点。仿真结果表明,容错检测算法校验后,主星检出率较传统方法提高10%,质心定位精度较平方加权质心法提高15%,基于复原星图的定姿误差小于20″。所提方法能有效提高临近空间高动态飞行环境下星光导航系统的适应能力,且对飞行器快速发射情况下初始位置不准引起的制导误差有较好的修正作用。Hypersonic vehicles operating in near-space environments are characterized by high-speed and highdynamic maneuvers.During flight,the high angular speeds can result in significant degradation of star sensor imagery,leading to a substantial decrease in the success rate of star pattern recognition.To address this challenge,an integrated framework based on deep learning,encompassing“detection-coarse localization-reconstruction-fine localization”,is proposed.Traditional star point extraction algorithms often exhibit large centroid extraction errors in high-dynamic environments.Therefore,a new approach is devised where the approximate location of star points is determined through a primary star pattern segmentation algorithm based on Faster R-CNN.This is succeeded by an end-to-end reconstruction network,utilizing a dense connected structure,to obtain precise centroid positions.Considering the extremely low signalto-noise ratio of real star images and their complex noise characteristics,a benchmark-star-assisted fault-tolerant detection algorithm is employed to validate the fine localization results,enabling the removal of misidentified stars and those with substantial localization errors.Simulation results indicate that the fault-tolerant detection algorithm achieves a 10%increase in the primary star detection rate,and the centroid localization accuracy is improved by 15%compared to the squared-weighted centroid method.Furthermore,the attitude determination accuracy based on the reconstructed star image is less than 20 arcseconds.The proposed method significantly enhances the adaptability of stellar navigation systems in high-dynamic near-space environments and provides substantial corrections to guidance errors arising from inaccurate initial positions during rapid launch scenarios.

关 键 词：端对端识别 星图拖尾 运动模糊 星图复原 临近空间高动态飞行器 

分 类 号：V249.322[航空宇航科学与技术—飞行器设计]