CASIA OpenIR  > 毕业生  > 博士学位论文
工业零件三维位姿检测、跟踪与装配方法研究
朱文俊
学位类型工学博士
导师乔红
2017-08-28
学位授予单位中国科学院研究生院
学位授予地点北京
关键词工业机器人 工业零件 自动装配 三维位姿检测与跟踪
摘要机器人技术在工业生产中的应用能够大大减轻劳动者工作强度,提高生产效率和产品质量,在汽车、机电、通用机械、铸造等工业领域,得到了越来越广泛的应用。随着工业自动化水平的不断发展,人们对机器人的各项功能也提出了越来越高的要求,特别是机器人的智能化水平。将视觉系统引入到机器人系统中能有效地提高机器人对环境的感知能力和智能化水平,其在工业零件的装配、搬运和分拣中具有重要的应用前景。工业零件的位置和姿态信息的获取,即工业零件的三维位姿检测,是机器人视觉系统应用领域中的核心技术和难点之一,也是目前制约机器人在定制化和个性化工业生产中应用推广的技术瓶颈之一。
工业环境中存在的噪声干扰,工业零件类型的多样性,以及工件堆放时的相互遮挡和投影变形等问题,对工业零件三维位姿的检测、跟踪与装配方法提出了新的挑战。现有方法,对于工业零件位姿的三维检测,存在着杂乱环境下识别率低、无法实现实时检测和跟踪、特殊位姿条件下检测错误率高等问题;对于装配存在着如何将机械约束和传感器信息有效融合,以提高装配的效率和精度等问题。本文针对上述实际问题,开展方法和应用研究,论文的主要研究工作和贡献总结如下:
(1)针对工业应用中,大多数的工件为无纹理且颜色单一,特征信息有限,难以实现准确鲁棒的位姿检测的问题,提出了基于离线建模和在线层级搜索的工件三维位姿检测方法。该方法利用工件的三维CAD模型和工业相机的内参数,离线生成层级匹配库;对工业相机实时采集到的图像,在线进行层级搜索,并对搜索到的目标物姿态进行位姿迭代优化,获得高精度的工件三维位姿。基于所提出的三维位姿检测方法,实现了单目视觉下机器人对三维工件的抓取。
(2)针对在杂乱的工业环境中,仅仅使用边缘特征来进行位姿测量,易造成误检测的问题,提出基于CAD模型和显著特征选择的工件三维位姿检测方法。该方法将工件的三维CAD模型和由该相机捕获的工作场景的系列图像作为输入,生成不同位姿下工件的视图库,以及包含目标物和工作场景的颜色、纹理等信息的显著特征库;在工件的位姿在线检测阶段,通过使用融合多特征的显著特征库来提高工业零件位姿测量的准确性和效率。
(3)针对基于视图的工件三维位姿测量方法检测速度慢,难以应用于对工件位姿的实时测量与跟踪这一问题,提出基于动态模型库的三维工件位姿跟踪方法。该方法主要包括静态视图库生成、动态视图库更新与选择,以及三维工件定位三个过程。方法通过对工件三维位姿的预测,实现动态视图库更新与选择,大大缩小了位姿搜索空间,提高了工件的三维位姿测量速度,成功实现单目视觉下三维工件位姿的实时测量与跟踪。
(4)针对如何将装配对象的机械约束信息引入到装配过程,与视觉传感器得到的位姿信息形成互补性优势,实现装配过程的定位-抓取-引导-插入的闭环控制的问题,设计构建了基于视觉引导和环境吸引域的曲柄轴-轴承装配系统。在视觉引导阶段,针对由不当的观测位姿,而引起的工件位姿检测错误问题,提出基于预分析与准确性估计模型的三维位姿检测方法。该方法采用预分析方法生成位姿检测准确性估计模型,进而对位姿测量的准确度进行判定,对于检测准确度低的情况,通过工业机器人的引导,改变相机检测角度,从而提高工件位姿检测的准确性。系统基于视觉引导和环境吸引域融合的方法,提出了曲柄轴-轴承装配策略,实现曲柄轴-轴承的高精度装配。
所提出的方法,利用搭建的工业机器人视觉系统实验平台,开展了不同工业零件的三维位姿检测、跟踪实验,同时成功实现了视觉引导下的曲柄轴-轴承装配,验证了本文所提出方法的有效性。
其他摘要Introducing robotic technology to industrial production can greatly reduce the work intensity of workers, and improve the production efficiency and product quality. Robotic technology has been more and more widely used in auto industry, mechanical and electrical industry, general machinery industry, casting industry and other industrial fields. With the development of industrial automation level, higher requirements for the function of the robot is also put forward, among which the intelligent level of the robot is an important content. The introduction of visual system into the robot system can greatly improve the robot’s ability to perceive the environment and the robot’s intelligent level. It has important application prospects in the fields of sorting, assembly and transportation of industrial parts. Acquisition of the pose and position information of industrial parts, i.e., industrial parts detection and localization, is a core technology in the application of robot vision system and one of the difficulties, and it is also one of the bottlenecks that restrict the promotion of robot in the customization and personalization in industrial production field.
New challenges arise for pose measuring and localization of industrial parts, because of the noise existed in the industrial environment, the diversity of types of industrial parts, and the occlusion and projection deformation of the work pieces when stacked together. The existing technologies for pose and position estimation of the work piece have some common problems, such as low recognition rate under cluttered environment, being unable to realize pose estimation and tracking in real time, high pose estimation error rate in some special pose situations, etc. Efforts on analytical methods and application researches are made to solve the above mentioned practical problems. Main work and contributions of the dissertation are summarized as follows:
 (1) In industrial applications, most of the work pieces have no texture and are of single color, and their feature informations are very limited. This makes detection errors easy to happen. To solve this problem, a pose estimation algorithm for work pieces based on off-line model construction and on-line hierarchical search is proposed. 3D work pieces grasp and manipulation is realized by introducing the vision system into the robot grasping system. The algorithm takes the 3D CAD model of the work piece as input, and a hierarchical template library is generated offline combining the internal parameters of the camera. Hierarchical searching strategy is conducted online to the images acquired by the camera, and precise pose of the recognized object is obtained through iterative optimization algorithm.
(2) For the problem that the detection error rate will increase when measuring the pose of the work piece using only edge features in clutter background. A method combining multi features to significantly improve the accuracy and efficiency of pose measuring is proposed. This method takes the CAD model of the work piece, internal parameters of the camera and a series of work scene images captured by the camera as input, and a 2D view  library is generated, also a salient features library is generated, which contains the color information of the scene and texture information of the work piece. In the online pose measuring phase of the work piece, the accuracy and efficiency of image matching are improved by using the salient feature library.
(3) For the problem of low pose estimation speed when using view based pose estimation method, which makes it difficult to realize real time pose estimation and tracking for the work pieces, a method for work piece pose estimation and tracking based on dynamic model libraries is proposed. The proposed method mainly consists three processes, i.e., the offline static global library generation process, the online dynamic local library updating and selection process, and the 3D work-piece localization process. The dynamic local library updating and selection is realized by pose estimation of the work pieces, and this greatly reduces the searching space, thus improves the pose measuring speed for the work piece. Monocular vision based real time pose measuring for the 3D work pieces is realized.
(4) A crank shaft bearing assembly system based on vision guide and environmental attraction field is designed and constructed to solve the problems of how realize close loop control for the localization-grasping-guidance-insertion during the assembly process by introducing the mechanical constraint information of the object to be assembled into the assembly to reach the complimentary advantage with the pose and position information gained by the vision sensor. High precision assembly is realized by using the proposed strategy with only the vision information. In the vision guide process, a pose measuring method based on preanalysis and accuracy estimation model is proposed to solve the problem of detection error caused by inappropriate observation pose. A pose measuring accuracy estimation model is constructed by preanalysis, then it is used to calculate the accuracy of the pose measuring result. For low accuracy pose measuring situation, the pose of the camera will be changed by changing the pose of the industrial robot, and thus raise the accuracy of the pose measuring. Using the proposed crank shaft bearing assembly strategy based on the environmental attraction field system, high precision assembly of the crank shaft and bearing is realized.
A platform of industrial robot with vision system is built. Experiments on pose measuring and tracking with several work pieces are conducted, and the crank shaft and bearing is successfully assembled by the guidance of the vision system.
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/20376
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
推荐引用方式
GB/T 7714
朱文俊. 工业零件三维位姿检测、跟踪与装配方法研究[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
朱文俊---博士学位论文.pdf(4973KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[朱文俊]的文章
百度学术
百度学术中相似的文章
[朱文俊]的文章
必应学术
必应学术中相似的文章
[朱文俊]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。