用于地球静止轨道目标的光学检测算法  

Optical Detection Algorithm for Geostationary Space Targets

作　　者：韩冰[1,2] 王晨希 翟智 刘乃金[3,4] HAN Bing;WANG Chen-xi;ZHAI Zhi;LIU Nai-jin(School of Future Technology,Xi’an Jiaotong University,Xi’an 710049,China;School of Mechanical Technology,Xi’an Jiaotong University,Xi’an 710049,China;Spatial Intelligent Manufacturing Research Center,Xi’an Jiaotong University,Xi’an 710049,China;China Academy of Space Technology,Beijing 100081,China)

机构地区：[1]西安交通大学未来技术学院,西安710049 [2]西安交通大学机械工程学院,西安710049 [3]西安交通大学空间智能制造研究中心,西安710049 [4]中国空间技术研究院,北京100081

出　　处：《空间碎片研究》2023年第1期1-10,共10页Space Debris Research

基　　金：国家自然科学基金(U22B2013,52105480)。

摘　　要：针对地球静止轨道(GEO)空间目标探测任务中目标特征薄弱、尺度小和定位精度要求高的问题,提出SFFRetinaNet(Shallow focus and FreeAnchor RetinaNet)算法。该算法针对空间目标特征提取不充分的问题,设计了一种聚焦浅层特征的残差网络结构,增强了网络对图像浅层特征的提取能力;引入了FreeAnchor检测器,将锚框匹配策略转化为极大似然估计问题进行优化,提高了目标检测框的定位精度;针对观测图像中目标样本数量匮乏、分辨率低及分布不均匀的问题,引入多分辨率融合的Copy-Paste数据增强方法,提高了算法的检测效果。SFF-RetinaNet算法在Kelvins SpotGEO挑战赛的数据集上进行了测试,mAP达到了71.28%,相较原算法提高了12.33%,算法检测速度提高了3fps,能够更好地应用于地球静止轨道空间目标检测任务。Aiming at the problems of weak target features,small scale and high positioning accuracy requirements in geostationary orbit(GEO)space target detection missions,the SFF-RetinaNet algorithm is proposed.Aiming at the problem of insufficient extraction of spatial target features,this algorithm designs a residual network structure focusing on shallow features,which improves the network’s ability to extract shallow features of images;introduces the FreeAnchor detector,transforms the anchor frame matching strategy into Optimized for the maximum likelihood estimation problem,improving the positioning accuracy of the target detection frame;in view of the lack of target samples in the observation image,low resolution and uneven distribution,the Copy-Paste data enhancement method of multi-resolution fusion is introduced,which improves the detection effect of the algorithm.The SFF-RetinaNet algorithm was tested on the data set of the Kelvins SpotGEO Challenge,and the mAP reached 71.28%,which is 12.33%higher than the original algorithm,and the detection speed of the algorithm has increased by 3fps,which can be better applied to space targets in geostationary orbit detection tasks.

关 键 词：地球静止轨道 目标检测 卷积神经网络 数据增强 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]