摘要
针对空间在轨服务等任务中的非合作目标部件的检测问题,因缺乏大规模带标签数据,提出一种基于无监督学习的空间非合作目标部件检测算法。首先利用基于文本提示的生成式模型生成类别可知的空间非合作目标的典型部件,然后将部件进行结构拼接,组成完整卫星模型,最后随机化卫星的位置与姿态,模拟其与观测星的相对状态,并基于光线投影模拟点云的遮挡效果,生成逼近深度相机实际采集的非完整点云数据及伪标签。利用这些生成的类别可知的伪标签对所提算法进行训练,并在开源卫星点云数据上进行测试。实验结果表明,所提算法在无监督条件下的平均精度均值(mAP)达到46.21%,相比基于类别不可知的聚类的无监督方法提高12.51百分点,验证了其在空间目标检测领域的有效性与潜力。
To address the challenge of detecting non-cooperative target components in space on-orbit servicing tasks where large-scale labeled data is lacking,this paper proposes an unsupervised learning-based detection algorithm for space non-cooperative target components.The proposed algorithm first employs a text-prompt-based generative model to generate typical components of non-cooperative space targets with class-aware categories.These components are then structurally assembled into complete satellite models.Subsequently,the position and orientation of the satellite are randomized to simulate relative states with the observing satellite.Ray casting is used to simulate occlusion effects on the point cloud,generating incomplete point cloud data and pseudo-labels that approximate the actual output of depth cameras.These class-aware pseudo-labels are used for training,and the algorithm is tested on open-source satellite point cloud data.Experimental results show that the proposed method achieves a mean average precision(mAP)of 46.21%under unsupervised conditions,outperforming clustering-based unsupervised methods with class-agnostic labels by 12.51 percentage points.This demonstrates its effectiveness and potential in the field of space target detection.
作者
姜子墨
闫浩东
朱振才
丁国鹏
Jiang Zimo;Yan Haodong;Zhu Zhencai;Ding Guopeng(Innovation Academy for Microsatellites,Chinese Academy of Sciences,Shanghai 201304,China;University of Chinese Academy of Sciences,Beijing 100049,China;School of Information Science and Technology,ShanghaiTech University,Shanghai 201210,China)
出处
《激光与光电子学进展》
北大核心
2025年第22期82-90,共9页
Laser & Optoelectronics Progress
关键词
机器视觉
无监督学习
三维目标检测
点云
空间非合作目标
生成式模型
machine vision
unsupervised learning
three-dimensional object detection
point cloud
spatial noncooperative target
generative model
