基于深度学习的空间目标部件智能分割与姿态估计  

Intelligent segmentation and pose estimation of space non-cooperative targets based on deep learning

作　　者：曹姝清 谭龙玉 罗建军[3] 尤浩楠 易建军[4] CAO Shuqing;TAN Longyu;LUO Jianjun;YOU Haonan;YI Jianjun(Shanghai Institute of Spaceflight Control Technology,Shanghai 201109,China;Shanghai Key Laboratory of Aerospace Intelligent Control Technology,Shanghai 201109,China;College of Astronautics,Northwestern Polytechnical University,Xi'an Shaanxi 710072,China;School of Mechanical and Power Engineering,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]上海航天控制技术研究所,上海201109 [2]上海市空间智能控制技术重点实验室,上海201109 [3]西北工业大学航天学院,陕西西安710072 [4]华东理工大学机械与动力工程学院,上海200237

出　　处：《航天工程大学学报》2024年第3期89-94,共6页

基　　金：国家自然基金航天企业联合基金重点资助项目(U20B2056)。

摘　　要：针对空间目标相对导航技术需求,提出一种基于深度学习的空间目标部件智能分割与姿态估计方法。通过虚实结合完成空间典型目标样本数据库的获取与增广构建,采用手动标注和自适应分割相结合实现目标部组件点云语义信息的快速标注。利用基于深度学习的PointNet++分层架构网络模型,完成对空间卫星5类典型部组件的三维识别与分割。并基于部组件智能三维分割结果,采用“整体建模,局部匹配”实现该部组件的位姿估计。算法无须对目标特定结构的特征提取与轮廓分割来实现对目标局部部组件的姿态估计,能有效克服复杂环境下特征提取慢且效果差导致姿态估计精度低的缺点,具有目标识别泛化性强、部组件分割速度快、姿态估计精度高等优点。For the technical requirements of relative measurement and navigation of space non-cooperative targets,an intelligent segmentation and pose estimation method of space non-cooperative target part components based on deep learning is proposed.Firstly,the target sample database is expanded and constructed by combining virtual simulation with real data acquisition Manual annotation and adaptive segmentation are used to quickly annotate the point cloud semantic information of the target component.Secondly,a PointNet++hierarchical network model based on deep learning is used to realize the three-dimensional recognition and segmentation of five typical components of space spin non-cooperative target at fixed point observation.Finally,based on the intelligent 3D segmentation results,the dynamic recognition and tracking for non-cooperative targets local specific components in the process of approaching measurement from far to near is completed and the realtime relative pose information for the local specific components is continuously estimated based on the iterative closest point method.The feasibility of the method is verified by building a ground experimental test system.The proposed algorithm does not require the feature extraction and contour tracking of the target specific structure to realize the local components pose estimation,which effectively overcomes the disadvantages of low pose estimation accuracy due to slow and ineffective feature extraction and contour tracking in complex environment and has the advantages of strong generalization of target recognition,fast component segmentation and high pose estimation accuracy.

关 键 词：空间目标 深度学习 三维分割 姿态估计 

分 类 号：V448.2[航空宇航科学与技术—飞行器设计] TP391.9[自动化与计算机技术—计算机应用技术]