基于特征学习的空间非合作目标单目视觉位姿测量研究

It’s proposed that the method of pose measurement of the non-cooperative target in space using monocular vision based on feature learning, and the strategy of “feature learning based on monocular vision” + “pose estimation based on feature corresponding ” is established. Firstly, the machine learning is melt into the FAST(Features from Accelerated Segment Test) and BRIEF(Binary Robust Independent Elementary Feature) to extract and match feature, which achieves feature accuracy and efficiency of extraction and matching because the feature matching problem skillfully converted to feature classification problem using the unsupervised machine learning; Then the method combined between fundamental matrix estimation and bundle adjustment is applied to establishing 3D feature library of target in image sequences, which solves reasonably the problem of maintain precision of 3D feature reconstruction with feature mismatching, and keep contradictions balance well between the complexity and the speed of reconstruction ; Finally, PPPIT (Pose from Orthography and Scaling with Iterations) and PROSAC(PROgressive random SAmple Consensus) are integrated to put forward the method of removing outliers of corresponding between 2-D image features and 3D feature library ,which improves the accuracy of pose measurement. Therefore, the study on pose measurement of the non-cooperative target using monocular vision based on feature learning will be Solid foundation in theory and practice for much space application including spacecraft RVD(rendezvous and docking) and robot capture.

本项目研究提出一种基于特征学习的空间非合作目标单目视觉位姿测量方法，并建立了“单目特征学习”+“特征对应测姿”策略。首先将机器学习思想创新融入到加速分割测试特征检测（FAST）和二元鲁棒独立基本特征（BRIEF）提取匹配过程中，将特征匹配问题巧妙转换为特征在无监督学习下的分类问题，达到特征提取匹配准确性、高效性；随后将基本矩阵估计和光束平差方法相结合用于序列图像的三维特征库建立，合理解决了误匹配下三维特征重建精度保持问题，平衡了重建速度和复杂度之间矛盾；最后提出平行投影的迭代测姿方法（PPPIT），并融合累进随机抽样一致（PROSAC）思想，提出图像二维特征与三维特征库之间2D-3D特征对应去野值方法，提高了位姿解算精度。基于特征学习的空间非合作目标单目视觉位姿测量方法研究将为航天器交会对接、机器人抓捕等空间应用打下坚实的理论基础。

本项目研究提出一种基于特征学习的空间非合作目标单目视觉位姿测量方法，并建立了“单目特征学习”+“特征对应测姿”策略。首先将机器学习思想创新融入到加速分割测试特征检测（FAST）和二元鲁棒独立基本特征（BRIEF）提取匹配过程中，将特征匹配问题巧妙转换为特征在无监督学习下的分类问题，达到特征提取匹配准确性、高效性；随后将基本矩阵估计和光束平差方法相结合用于序列图像的三维特征库建立，合理解决了误匹配下三维特征重建精度保持问题，平衡了重建速度和复杂度之间矛盾；最后提出平行投影的迭代测姿方法（PPPIT），并融合累进随机抽样一致（PROSAC）思想，提出图像二维特征与三维特征库之间2D-3D特征对应去野值方法，提高了位姿解算精度。基于特征学习的空间非合作目标单目视觉位姿测量方法研究将为航天器交会对接、机器人抓捕等空间应用打下坚实的理论基础。

内容获取失败，请点击重试