Multi-Stage Star Image Identification Method of Three Field-of-View Star Sensor

Bin Gou,Huibin Wang,Yongmei Cheng,Chengyuan Liu,Mingyan Zhao

doi:10.1051/jnwpu/20193730541

Abstract

To improve the low efficiency and low navigation star identification rate of existing star image identification methods for three field-of-view (FOV) star sensor, a multi-stage star image identification method is proposed. Firstly, the generalized regression neural network which has only one adjustable parameter, is used to identify the star images in each field-of-view. Secondly, the star angular distance saved in the navigation star database is used to verify the identification results, and then the optical directions of the three FOVs are calculated by using the correctly identified navigation stars. Thirdly, the optical directions are utilized to auxiliary correct the unidentified and erroneous identified navigation stars. Finally, the high-accuracy probe attitude is estimated by using the correctly identified navigation stars in the three FOVs. The simulation results show that the identification rates of the experimental samples is of 98.9% when the standard deviation of star centroid positioning error increases to 0.07 pixels, but the identification time is only of 8.464 5 ms. Meanwhile, since the three field-of-view star sensor captures the more dispersed navigation stars, the probe attitude accuracy of yaw, pitch and roll angles by using the present method is improved evidently, which is of 1.205 8″, 1.086 7″, and 1.201 8″, respectively.

Highlights

The star angular distance saved in the navigation star database is used to verify the identification results, and the optical directions of the three FOVs are calculated by using the cor⁃ rectly identified navigation stars
The simulation results show that the identification rates of the experi⁃ mental samples is of 98.9% when the standard deviation of star centroid positioning error increases to 0.07 pixels, but the identification time is only of 8.464 5 ms
Since the three field⁃of⁃view star sensor captures the more dispersed navigation stars, the probe attitude accuracy of yaw, pitch and roll angles by using the present method is improved which is of 1.205 8′′, 1.086 7′′, and 1.201 8′′, respectively

Summary

Introduction

西北工业大学学报 Journal of Northwestern Polytechnical University https: / / doi.org / 10.1051 / jnwpu / 20193730541 定位误差由 1 增至 0.07 个像素。图 4 统计了本文算法、三角形匹配算法( 全天球识别) [11] 、P 值向量算法[11] 、RBF[2] 以及 BP [3] 星图识别算法在不同导航星质心定位误差下的识别率以及平均识别时间。其中,RBF、BP 神经网络的训练集与 GRNN 相同。由图 4 可以看出,神经网络类方法具有相似的星图识别时间, 且明显优于其他方法。但本文的 GRNN 星图识别算法在网络训练之初就考虑了星点位置误差干扰,增加了网络的容错能力。当星点位置误差标准差为 0.07 像素时,本文算法在 1 000 组实验样本中有 920 组 3 个视场被同时正确识别。 2.2 视轴指向辅助视场间星图校正结果分析本文提出了一种面向三视场星敏感器多级星图识别算法,第一阶段利用神经网络进行视场内星图识别,第二阶段以星库中存储的星间角距信息检验导航星识别结果,当正确识别的视场数小于 3 时,利用已正确识别的导航星信息计算星敏感器的 3 个视轴指向;第三阶段利用视轴指向辅助未识别与识别错误的导航星完成识别与校正。与文献[7] 相比, 本算法的视场内星图识别鲁棒性与识别时间具有明显优势。当星点质心定位误差标准差为 0.07 像素时, 本文算法对实验样本的识别正确率仍保持 98.9%,识别时间仅为 8.464 5 ms。三视场星图的高识别率与合理的星点分布也使得飞行器偏航、俯仰、滚转姿态精度分别提升为 1.

Results

Conclusion