面向大口径器件装配的三维检测与位姿估计方法研究

CASIA OpenIR > 毕业生 > 博士学位论文

	面向大口径器件装配的三维检测与位姿估计方法研究
	覃政科
	2017-05-25
学位类型	工学博士
中文摘要	大口径器件装配是航空航天、船舶、汽车、大科学装置等制造过程的重要环节之一，随着我国智能制造战略的推进，对大口径器件的装配质量要求越来越高，装配技术也更趋向于自动化和智能化方向发展，对装配对象和装配位置的精确三维检测和位姿估计是实现大口径器件自动化装配的重要前提。由于装配对象尺寸大、作业跨度大、精度要求高，给大口径器件装配过程的三维检测和位姿估计带来了挑战：（1）装配过程作业空间跨度大，从米级别到几十毫米级别，动态范围大；（2）装配对象尺寸大（百毫米甚至米级），而装配精度要求高（毫米或亚毫米级）。本文针对大口径器件装配过程三维检测与位姿估计存在的瓶颈问题开展研究，对于测量范围的问题，本文论将从米级到亚毫米级别的测量空间进行划分，得到远距离引导和近距离对准子问题。对于精度问题，对远距离引导提出基于链接模型与梯度方向三维模板匹配的目标检测与六自由度位姿估计方法，对近距离对准过程中提出了基于多参考点的表面跟踪方法以及基于激光位移传感器的对准方法。更详细地，本论文主要内容和贡献包括：（1）针对大口径器件装配中对三维检测与位姿估计的实时性要求，提出基于链接模型的多边形快速三维检测与位姿估计方法。在图像经过去阴影和平滑的预处理之后，提取图像中的直线段，根据线段的端点之间的几何关系选择邻边，构建链接模型，最后通过深度优先法搜索链接模型中的闭合线段搜索多边形，依据多边形顶点之间的几何关系，计算出三维空间内的平面多边形和非平面多边形的位姿。通过对大口径器件的两个同心矩形进行检测，完成对大口径器件的快速检测和位姿估计。相对于一般的匹配方法，通过直线检测方式避免了在巨大的像素空间的搜索，将检测搜索问题复杂度变换到直线段条数的数量级，达到快速检测与初步位姿估计的目的。此外，针对于大口径器件的检测和位姿估计中的去阴影问题，本文还提出一种检测阴影的方法，基于几何信息和纹理信息，以“求位姿-去阴影进行迭代”模式，去除图像中和物体有着相似的轮廓的阴影，提高器件位姿估计精度。（2）针对大口径器件装配中的远距离的实时性、高精度三维位姿估计要求，以及链接模型方法在精度上的不足，提出基于链接模型和梯度方向三维模板匹配的三维检测与位姿估计方法。基于链接模型检测图像中的所有多边形，并且通过五点交比不变量筛选出表达检测目标的多边形，快速地得到目标的初步六自由度位姿。然后通过梯度方向三维模板匹配方法获取目标精确的六自由度位姿，其中包括构建目标投影模板库，搜索最相近模板以及三维模型的动态对准。使用链接模型可以快速得到大口径器件的初步位姿，限定梯度三维模板匹配的搜索空间。使用梯度方向三维模板匹配的方法可以在达到一个很高的位姿估计精度。该方法有效利用了两者的优势，在检测速度和检测精度上达到很好的平衡，对于大口径器件的装配更加实用。（3）针对于大口径器件装配过程中的近距离连续对准问题，提出基于多参考点与单应性矩阵的表面位置跟踪方法。基于 Shi-Tomasi 角点检测方法提取图像中的参考点，并且对参考点赋予梯度方向直方图特征，基于最近邻匹配方法对参考点进行匹配，基于相邻两帧图像的单应性矩阵估计目标跟踪点。通过该跟踪方法，完成了在近距离大口径器件对准过程中的图像伺服目标点实时、准确的跟踪。为大口径器件的对准过程提供两个平移自由度以及一个旋转自由度的精确反馈。（4）在上述算法研究的基础上，针对大口径器件大范围、高精度装配任务，提出基于视觉-激光位移传感器的大口径器件三维检测、位姿估计与装配方法。通过远近两个相机和三个激光位移传感器完成对大口径器件装配过程中由远及近的测量范围的覆盖。基于三个串行装配阶段：远距离视觉引导，近距离视觉-位移传感器对准以及自动插入，完成大口器件由远及近、完全自动化的精确安装。其中，链接模型-梯度方向三维模板匹配方法保证了远距离引导的实时性与精确性。多参考点跟踪表面跟踪方法与基于激光位移传感器测量方法保证了近距离对准与插入的精确性。基于链接模型与梯度方向三维模板匹配方法检测大口径器件的基础上，为装配系统提出了一种基于期望安装位姿的相机-机器人手眼标定方法，只需要使用梯度方向三维模板匹配的方法从不同姿态对目标进行若干次检测定位，以及完成一次期望安装即可，节省标定工作量。论文的研究成果应用于大口径光学器件的自动化装配过程，为提高我国大科学装置的建设质量和效率提供了技术支撑，相关的成果也可推广应用于航空航天、汽车制造等领域。
英文摘要	In the area of manufacture for aerospace, aviation, automobile and large experiment equipment, the automatic assembly for large scale components is an important requirement. As the large equipments require higher quality of automatic assembly for large scale components, the assembly technology is developed for more automatic and higher precision. The efficient object detection and precise pose estimation is an important precondition for large scale components assembly. The complete assembly for large scale object assembly has two main problems: (1)The assembly range is covered from meters to dozens of millimeters which is relatively large, as a consequence, a single sensor based method is hard to handle the whole range. (2)The assembly precision requirement for the objects with the scale of hundreds of millimeters (or even meters) is usually sub-millimeters, which means the relative assembly precision is relatively high. This dissertation is focus on the 3D detection and pose estimation for large scale component assembly, especially the vision based methods. For the large measurement range problem, this dissertation separates the measurement space from meters to millimeters to two sub-problems: the long distance guidance and close range alignment. For the precision problem, this dissertation proposes the link line model and gradient orientation 3D template based method for long distance guidance, the multiple reference points based surface position tracking method and laser displacement sensor based method for close range alignment. To go further, the main work and contributions are as follows: (1) A link line model based polygon detection method is proposed, for the realtime detection requirement of large scale components assembly. After the preprocess of shadow removement and Gaussian smooth, the line segments of the input image are extracted by the LSD method. The neighbors of each line segment are selected base on their geometric relationship, to build the link line model. At last link line model is traversed base on deep first strategy for searching the polygons. The poses of the planar and non-planar polygons in 3D space are estimated base on the geometric information between the vertices. The large scale component is detected and its pose is estimated, by detecting the polygons on its surface. With respect to matching methods, this line segment based method avoids the huge pixel searching space, instead, it turns the complexity to the order of the line segment number. Base on this method the large scale component is detected quickly with an acceptable precision. For the preprocess, a shadow detection method during object detection and pose estimation is developed, for removing the shadow which has similar contour to the object, base on geometry and texture information. (2) To meet the high pose estimation precision and real-time requirement for large scale components in a relative long range, this dissertation proposes the link line model - gradient orientation 3D template based method. The polygons of the input image are detected base on the link line model, the component’s polygon is extracted with the co-planar invariants and the component’s pose is attained quickly. And then the precise pose is estimated by the gradient orientation 3D template method, including the creation the off-line object projection templates library, searching for the most similar template and aligning the 3D model to the input image dynamically. By applying the link line model method, the large scale component is detected and its coarse pose is estimated quickly. After the search range of the gradient orientation 3D template method is limited, the pose of the large scale object can be attained both precisely and quickly. This method takes the advantages of both of them and achieves a relative good balance between precision and time consuming, which is more practical for large scale components assembly. (3) To solve the alignment problem in the close range for large scale object assembly, a surface location tracking method base on multiple reference points and homography matrix. The reference points are detected by Shi-Tomasi corner method, and these reference points are assign the feature description of gradient orientation histogram. And then the reference points between two frames are matched by the nearest neighbor method, and at last the object location point is calculated by homography matrix. According to this tracking method, the objective point during the visual servoing process is tracked precisely in real-time. Two translate components and one rotation component feed back are provided for the close range alignment process. (4) To complete the far-to-close, precisely and real-time assembly task for large scale components, a vision - laser displacement sensors based method is proposed. The far-to-close assembly measurement range is covered by two cameras with different working range, and three laser displacement sensors. The large scale component is assembled completely automatic from far to close, base on three stages: the long distance visual guidance, the close range visual and displacement alignment and the automatic insertion. link line model ensures the real-time and precision requirement during the long distance guidance. The multiple reference points surface location tracking and laser displacement based methods ensure the precision of the close range alignment. Base on our link line model - gradient orientation 3D template matching method, a camera robot hand eye parameters calibration method is proposed. This method calculates the hand eye parameters base on the objective installation pose, which is attained by precisely manually installation. This calibration method does not need to sample a large number of pattern images in different robot poses. Instead, it only requires several detection and pose estimation processes and one manually assembly, which makes it a precise and easily applied method. The methods proposed by this dissertation is applied in the automatic assembly task for large scale optical components, provide the technology supports of the quality and efficiency for large scale scientific experiment equipments’ constructions. The proposed methods can also be introduced to other areas, such as aerospace, aviation and automobile manufactures.
关键词	大口径器件装配视觉检测六自由度位姿估计视觉跟踪
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/14758
专题	毕业生_博士学位论文
作者单位	中国科学院自动化研究所
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	覃政科. 面向大口径器件装配的三维检测与位姿估计方法研究[D]. 北京. 中国科学院研究生院,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
面向大口径器件装配的三维检测与位姿估计方（13287KB）	学位论文		限制开放	CC BY-NC-SA