基于残差自注意力和分离集合匹配的高效端到端航天器组件检测

Efficient End-to-End Spacecraft Component Detection Based on Residual Self-attention and Separated Set Matching

随着我国航天技术的迅猛发展,各种航天器相继发射,然而航天器在运行时将受到辐射、温度变化等不可控因素的影响,这会导致地面站无法精确测量和定位航天器的位置与姿态,从而对通信和航天器之间的对接或抓捕等空间在轨服务产生影响。为了解决上述问题,首先对包含检测、分割与部件识别的航天器数据集SDDSP中的部件进行人工标注,该数据集共包含3117张航天器图片,标注后得到11001个检测目标;然后提出一种空间在轨服务中基于残差自注意力(RS)和分离集合匹配(SSM)的高效端到端航天器组件检测模型,该模型在Sparse DETR模型的基础上引入残差自注意力机制解决了稀疏标记(token)导致的收敛速度降低并影响模型预测精度的问题,引入分离集合匹配机制解决了二分匹配过程中可能出现的不稳定性现象。实验结果表明,在SDDSP数据集上,该模型的平均精确率(AP)和收敛速度相比于基线DETR模型提升了17.9个百分点和10倍,相比于Sparse DETR模型提升了3.1个百分点和20%。

The rapid development of space technology in China has led to a multitude of spacecraft launches.However,these spacecraft are expected to experience the influence of uncontrollable factors such as radiation and temperature changes during operation.These changes may impede the accurate measurement of spacecraft positions and behaviors by ground stations,thereby impacting on-orbit services such as communications and docking,as well as grappling between spacecraft.To solve these problems,the present study first annotates the SDDSP spacecraft dataset which encompasses detection,segmentation,and component recognition with 3117 spacecraft images and 11001 detection targets.An efficient end-to-end spacecraft component detection model is then proposed based on Residual Self-attention(RS)and Separated Set Matching(SSM)in space on-orbit services.The RS mechanism is introduced on the basis of the Sparse DEtection TRansformer(DETR)model to solve the problem of sparse tokens,which slows convergence and degrades the prediction accuracy of the model.Furthermore,SSM is deployed to address the phenomenon of instability that may occur in the process of dichotomous matching.The experimental results show that the Average Precision(AP)and convergence speed of the model are improved by 17.9 percentage points and 10 times,respectively,compared with those of the baseline DETR model,as well as 3.1 percentage points and 20%,respectively,compared with those of the Sparse DETR model.