【作者】 唐永鹤；

【导师】 卢焕章；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2012， 博士

【摘要】 空间非合作目标三维模型获取是在轨服务的关键技术之一，在太空领域有着广阔的应用前景。本文以对空间非合作目标进行绕飞、伴飞为背景，研究基于序列图像的空间非合作目标三维结构重建问题，主要围绕空间平台相机在线标定、多种变换条件下的快速图像匹配和快速精确鲁棒的基本矩阵估计三项关键技术进行研究。为了自动提取场景几何约束信息，对边缘检测算法进行了深入研究，提出了一种基于多结构元素复合滤波的形态学边缘检测算法。算法不仅能够有效检测图像的边缘，还具有抗噪性强、处理速度快的优点。为了降低场景约束条件，扩展空间平台相机在线标定的应用范围，提出了一种基于平行四边形相似不变量的相机自标定方法。同时，为了校正因相机畸变而产生的成像误差，提出了一种基于直线射影不变性的径向畸变参数估计算法。该标定方法不仅有效降低了场景约束条件，而且具有简单、灵活的特点。在空间平台硬件计算资源有限的情况下，为了实现存在多种变化的图像间的快速匹配，首先提出了一种Harris角点检测硬件加速方法，并将其应用到高斯金字塔图像上，实现多尺度Harris角点检测硬件加速，采用内嵌并行运算的流水线结构，提高算法的实时处理能力，减少硬件资源消耗。然后，提出了一种基于SCCH的局部特征描述算法，具有计算简单、独特性好、鲁棒性强的特点。最后，将二者结合在一起，实现多种变换条件下的快速图像匹配。为了快速、鲁棒、精确地估计基本矩阵，提出了应用序贯相似检测的快速鲁棒基本矩阵估计算法。算法引入SSDA加速策略提高其处理速度，用M估计算法优化初始内点集，降低估计余差较大内点的影响，提高算法的估计精度和鲁棒性。在本文研究成果的基础上设计了一种三维重建实验系统，为建立实际的基于序列图像空间非合作目标三维重建系统奠定了基础。实验结果表明，该实验系统是有效的和可行的。更多还原

【Abstract】 Three-dimensional (3D) model acquisition of non-cooperative space target is oneof the key technologies for on-orbit servicing, and has broad application prospects in thespace. Under the background that flying around or along with the non-cooperative spacetarget, this dissertation reseaches on3D structure reconstruction technologies fromimage sequences of non-cooperative space target. Three key technologies for3Dreconstruction are mainly investigated, including on-line camera calibration on spaceplatform, high-speed image matching with multiple transformations between images, aswell as fast, accurate and robust fundamental matrix estimation.To extract the geometric constraint information automatically, edge detectionalgorithm is studied, and a morphological edge detection algorithm based onmulti-strucure elements compound filter is presented. The proposed algorithm performsbetter in edge detection, noise immunity and speed. In order to lower the sceneconstraint requirements and extend the application scope, a camera calibration methodbased on parallelogram similarity invariants is presented. The first two terms radialdistortion parameters are also estimated according to the fact that straight lines in thescence are projected as straight lines in the image. The proposed algorithm not onlyreduces the restriction of sence constraint effectively, but also is simple and flexible.With the hardware computing resources limited, to match those images withmultiple transformations fastly, a hardware-accelerated method for Harris cornerdetector is developed firstly, and it is applied to the Gaussian pyramid images to achievehardware acceleration for multi-scale Harris corner detector. The method uses pipelinearchitecture with internal parallel operation to improve the real-time performance andreduce the hardware resources consumption. Secondly, a local feature descriptionalgorithm based on signed contrast context histogram (SCCH) is proposed, which issimple, distinctive and robust. Finally, the above two algorithm are integrated to matchimages fast under multiple transformations.To estimate the fundamental matrix rapidly, robustly and accurately, a fundamentalmatrix estimation method based on sequential similarity detection algorithm (SSDA) ispresented. SSDA is introduced into the maximum a posteriori sample consensus(MAPSAC) to increase the speed. The initial inliers obtained by improved MAPSACare optimized with M-estimator. Those inliers with larger residual errors are removed toenhance the precision and improve the robustness of the algorithm.A3D reconstruction system is designed based on the results of this study, whichlays a foundation for establishing actual3D reconstruction system for non-cooperativespace target using image sequences. The experimental results show that theexperimental system is effective and feasible.更多还原