【作者】 梁志敏；

【导师】 吴林；

【作者基本信息】 哈尔滨工业大学 ， 材料加工工程， 2008， 博士

【摘要】 核电站的维修、空间结构的建造以及海洋工程的建设大量用到了焊接技术,机器人遥控焊接技术是代替人进入这类危险、极限环境中执行焊接操作的最佳选择。在非结构化环境的遥控焊接中,通过机器人任务空间的三维重建,发挥机器人自动化功能可以避免对操作者的过分依赖、提高系统的效率和安全性。此外,焊接机器人全自主焊接的发展也需要提高机器人感知环境的能力。本文基于立体视觉技术,针对遥控焊接机器人任务空间的三维重建进行了深入研究,实现了无纹理焊接环境的三维建模。本文建立了用于遥控焊接机器人任务空间三维重建的立体视觉系统,并完成了系统标定和极线校正。系统标定建立了机器人坐标系中空间点与摄像机图像上像素点的映射关系。极线校正算法通过对图像进行重采样,将对应极线变换到图像扫描线上,降低了立体匹配算法复杂度。焊接场景中纹理的缺乏导致立体匹配尤为困难,本文采用主动光源向场景添加纹理,采用基于单图像对的特征匹配和基于图像序列的时空立体视觉算法完成立体匹配。在单图像对的立体匹配中,使用角点特征和线特征两种特征。采用亚像素检测算法提取角点特征,以人机交互或者自动匹配算法完成了外围设备和简单工件图像对的立体匹配。本文提出一种焊缝线特征检测算法,用于焊缝图像的立体匹配。该算法通过底帽变换形态学预处理,突出焊缝区域,用Canny边缘检测和闭操作获得初始的像素级焊缝边缘点;依据焊缝的连续性和宽度有限对初始边缘点进行筛选,剔除错误点;然后用三次平滑样条拟合获得了光滑的亚像素精度的焊缝边缘点。在极线约束下完成焊缝特征的立体匹配。本文设计了一种锯齿式多级对偶灰度光源序列向场景中添加纹理,并提出了归一化SSSD匹配算法,对时空立体视觉技术进行改进,解决了焊接无纹理环境的立体匹配问题。归一化SSSD匹配算法通过对不同时刻图像对SSD的归一化处理提高了匹配精度。采用曲线拟合的方法求解亚像素视差,通过左右一致性检验和图像序列的灰度跟踪,去除了视差图中的错误区域,并引入了各向异性扩散滤波对视差图进行保留深度突变的平滑处理。本文进一步对从视差图或者点云数据获得任务空间的三维模型进行了研究。根据视差图的特点可以直接从视差图生成场景的三角网格模型;另外,通过视差图的分割和曲面拟合也可获得焊接工件的标准几何模型。本文改进了USF距离图平面分割算法,增加了区域合并步骤,使其能够用于含有圆柱面的视差图的分割。对于分割后的属于不同表面的点云,通过曲面拟合的算法建立模型,这里对平面、一般二次曲面和圆柱面拟合进行了研究。其中,平面拟合采用特征向量分解法,一般二次曲面采用正则化最小二乘法拟合,而圆柱面拟合提出一种主成分分析法与非线性最小二乘迭代结合的新算法。针对焊缝,采用NURBS曲线拟合算法从焊缝点云数据建立了焊缝曲线模型。针对典型的焊接任务空间,进行了外围设备、焊接工件和焊缝的三维重建实验,并对系统标定误差和重建结果误差进行了分析。标定的绝对误差在X、Y、Z方向上绝对值的平均数为[0.26 0.18 0.74]mm,最大误差的绝对值为[0.42 0.40 2.15]mm。外围设备顶点的绝对定位误差为6.53mm;在焊接工件的重建实验中,平面工件拟合误差为0.92mm,边长平均偏差3.43mm;马鞍形工件的拟合误差为1.35mm,高度和圆柱半径平均偏差2.25mm。焊缝的重建结果中,S型焊缝的重建焊缝与真实测量数据的平均距离误差为1.10mm,圆柱对接焊缝的平均距离误差为1.33mm。重建实验结果和误差实验结果显示,本文基于立体视觉的三维重建算法可以克服焊接场景无纹理的缺点,获得精度较高的重建结果,能够满足遥控机器人任务空间的建模需要。更多还原

【Abstract】 Welding technology would be widely used in maintenance of nuclear plants, construction and repairing operations of underwater structures and outer space aircrafts. Telerobotic welding has become the best choice to perform welding tasks in those hazardous or extreme environments instead of human operator. When performing the weld task in unstructured environment, automatic robot functions can be employed by the recontruction of robot’s task space to avoid depending too much on human operator and to improve the efficiency and security. Furthermore, according to the development need of the completely autonomous welding, the ability to perceive envirionment of weld robot should be improved. The thesis investigates 3D reconstruction of the welding robot’s task space base on stereo vision, and 3D modeling of untextured weld environment is implemented.Stereo vision system for 3D reconstruction of telerobotic’s task space is set up. The system is calibrated and epipolar rectification is applied to the image pair. By system calibration, the mapping between the point in the robot coordinate and the pixel in the image is built. Epipolar rectification resamples the image and transforms epipolar line to image scanline, so that the complex of stereo matching algorithm is reduced.Lack of texture in weld scene makes stereo matching harder. Active illumination is applied to add artificial texture into the scene, and both feature matching of singe image pair and image sequence matching of spacetime stereo are investigated to accomplish stereo matching.In single image pair matching, features of corners and lines are used. Subpixel corner detection algorithm is used to detect the corner features. Man-machine interaction and autonomously corner matching are used for image pairs of peripheral equipment and simple workpiece respectively. A new line feature detection algorithm is presented to accomplish stereo matching of weld seam. The image is preprocessed by bottom hat transformation of morphology to enhance weld seam section and initial closed pixel level edge points of the weld seam are obtained by Canny detector and close operation. According to the continuity and limited width of weld seam, these edge points are filtered to eliminate wrong data. Then the filtered data is fitted to cubic smooth spline and the subpixel weld edge features are obtained. These subpixel feature points of weld seam are matched by epipolar constraint.A multi-level dual“saw-tooth”stripe gray light patterns are designed to add artificial texture into scene and a normalized SSSD matching algorithm is presented to revise spacetime stereo to accomplish the matching of images for the untextured weld environment. The normalized SSSD algorithm improves precision by normalizing the SSD of image pairs at different times. Subpixel disparity is calculated by fitting a second-degree curve to the SSSD values. By intensity tracking of the image sequence and left right consistant check, wrong regions are removed in the disparity map and anisotropy diffusion filtering is introduced to smooth disparity map while preserving depth discontinuity.The problem of creating 3D model of task space from disparity map or point clouds is also studied. The scene’s triangulation mesh model can be directly created from disparity map. Besides, standard geometry model of weld workpiece can be obtained by disparity map segmentation and surface fitting. The USF range image plane segmentation algorithm is revised to segment disparity map containing cylinder surface via an additional step of region merging. The segmented point clouds, which belong to different surfaces, are then fitted to different surface models. The fitting algorithms for plane, general quadric and cylinder are investigated. Eigenvector method is used for plane fitting and regulation least squares method is used for quadric fitting. For cylinder fitting, a new algorithm of collaboration of principal component analysis and nonlinear least-squares algorithm is presented. For weld seam, NURBS curve fitting is used to create model from point cloud of weld seamFor typical weld task space, the reconstruction experiments for peripheral equipment, weld workpiece and weld seam are carried out. The errors of system calibration and reconstruction are analyzed. The mean of absolute value of the calibration’s absolute error along X、Y、Z axes are [0.26 0.18 0.74]mm, respectively. Mean absolute position error for corner of peripheral equipment is 6.53mm. The fitting error for plane workpiece is 0.92mm, and the mean error of the boundary is 3.43mm. For saddle-shape workpiece, the mean fitting error is 1.35mm, and mean error for cylinders’heights and radius is 2.25mm. In reconstruction results of weld seam, the mean distance of fitted curve to the measured true datas for S-weld seam is 1.10mm, and the mean distance for cylinder butt joint weld seam is 1.33mm. The reconstruction results and errors show that the 3D reconstruction algorithm based on stereo vision can overcome the problem of lack of texture in welding scene, relative high precision can be obtained, and the algorithm can satisfy the need of creating model of task space for telerobotic welding.更多还原