【作者】 陈四杰；

【导师】 顾寄南；

【作者基本信息】 江苏大学 ， 机械设计及理论， 2013， 博士

【摘要】 具有物体识别能力是机器人具有智能性的一个重要标志。智能机器人通过感知目标物体的位置和形状,才能实现抓取或跟踪等任务。形状信息是一种高层视觉特征,通过形状信息特征可以让机器人实现对物体形状的理解。如何让机器人将目标图像从自然图像中抽象并表示出来,是计算机视觉的一个核心。关于纯形状的分析研究虽然取得了大量成果,但是受噪声、形状建模等因素影响,实际应用却很少。本文对服务于数控机床群协同加工的上下料机器人所抓取的工件、卡盘等目标物体的识别、定位进行了研究,受到格式塔心理认知机制的启发,采用了自底向上的特征提取和自顶向下的形状指导相结合混合策略,解决了工件和卡盘识别定位中的关键问题。论文的研究工作及取得的成果如下：(1)针对目标物体和背景混合在一起的自然图像,提出了一种基于格式塔心理感知理论的高层视觉引导下的不连续轮廓段的识别方法。该算法提出了局部形状片段描述子,并建立了形状片段相似度评价函数,确定片段与模型间的从属概率,增强了局部形状片段对模型的作用,提高了拟合过程的收敛速度。对不连续轮廓链,在图模型基础上对不连续轮廓链进行“完形”操作,以寻找虚实交错的最小代价的简单环将高次问题转化为优化问题,并将计算复杂度降低到多项式级。通过邻域投票机制,对背景噪声边缘进行剔除和抑制,剔除异值轮廓后,减少了数据计算量,提高了定位精度。通过对物体轮廓的曲线拟合,准确的得到了上下料的空间位置信息。(2)为了对毛坯工件进行准确选择,保证最小的加工余量,提出了基于修正的贝塞尔点扩散函数的亚像素边缘检测方法。根据初步定位边缘的灰度分布特征,采用最小二乘法拟合原理进行亚像素定位。拟合函数采用了修正贝塞尔点扩散函数。修正后的贝塞尔点扩散函数吸取了高斯点扩散函数的优点,引入可变参数对贝塞尔点扩散函数进行一次修正和二次修正,提高了的贝塞尔点扩散函数灵活性和精确性,其实验结果表明修正后的贝塞尔拟合比高斯拟合具有更高的精度,而且一次修正速度较快,满足了对于边缘检测的实时性要求。(3)研究了摄像机标定的主要方法,分析摄像机标定产生误差的主要原因,进行了误差纠正。在实验数据分析的基础上,总结了基线距、焦距、物距与标定精度的非线性关系,绘制了标定误差变化的曲线图。通过摄像机标定的内外参数和基线距的选择优化,使双目视觉标定的精度和稳定性得到提高。(4)在图像的降噪处理方面,提出了一种保留图像边缘细节特征的平滑滤波方法。该方法构造了基于九邻域的由五边形和六边形组成的窗口掩模,该掩模设计具有一个优势,当像素点位于一个形状复杂邻域的锐角处,该掩模也能找一个与边缘形状基本相符的均匀的邻域,这样在去除噪声的同时也尽可能的保留了边缘细节。上述研究工作为面向数控机床群的上下料机器人的工程实践奠定了理论基础,并且对提高加工制造业的自动化水平,增强计算机视觉技术在工业领域的适用性具有重要意义。更多还原

【Abstract】 The robot with object recognition is an important symbol of intelligence. Intelligent robot can realize the capture or tracking task through the position and shape of object perception. The shape information is a kind of high-level visual features, analysis on the characteristics of shape information can make robot realize the understanding of the object shape. How to make robots represent and abstract shape information of the aim object from natural images is a core of computer vision. A lot of achievements have been made in analysis and research on pure shape, but a few are rarely used in practice, influenced by the noise and shape model.In this paper, mainly research on positioning and objects recognition such as robot grapping workpiece, chuck, is about the robot special used in the service of NC Machine Group. Enlightened by Gestalt psychology cognitive theory, the paper puts forward the hybrid strategy, which combines the feature extraction of bottom-up and top-down the shape of the guidance, and solve the key problems in the workpiece and chuck identification and location.(1) According to the target object and the background are blended in the natural images, this paper proposes a high-level vision based on the theory of the Gestalt perception under the guide of discrete contour recognition method. The algorithm presents local shape fragment descriptors, and establishes the shape similarity evaluation function to determine the segment between the model and subordinate probability, enhances the role of local shape fragments on the model, improves the convergence speed of the fitting process. For the discontinuous contour chain,"Gestalt" operation on the discontinuous contour in figure on the basis of the model, finds a minimum cost simple ring. The high order problem can be converted to optimization problems, and the computational complexity is reduced to polynomial level. Through the neighborhood voting mechanism, the background noise is eliminated and inhibited. After excluding abnormal value profile reduced the data computation, the positioning precision is improved. The accurate space location information about feeding NC machine is got by the curve fitting of the contour of the object.(2) In order to select the blank workpiece accurately, and ensure the minimum machining allowance, this paper puts forward the sub-pixel edge detection methods based on modified Bezier point spread function. According to the gray level distribution of the initial edge localization, the subpixel is located by the quafric curve. The quafric curve uses the modified Bezier point spread function. Revised Bezier point spread function takes the advantages of Gauss point spread function. The introduction of variable parameters on the Bezier point spread function for a correction and the second amendment, improves the flexibility and accuracy of Bezier point spread function, the experimental results show that the revised Bezier fitting has a higher accuracy than Gauss fitting, and a correction is faster, to meet the real-time requirements for edge detection.(3) Research on the methods of camera calibration, and analyze the main cause of the camera calibration error and complete the error correction. Based on the analysis of experimental data, summarizes nonlinear relation between calibration accuracy and the baseline distance, focal length, distance; then draw the curve of calibration error change. Through the internal and external parameters of camera calibration and optimization choice of baseline distance, binocular calibration precision and stability is improved.(4) In the denoising image, this paper proposes a smooth filtering method of retention characteristics of image edge details. This method constructs window mask based on nine neighborhoods with pentagons and hexagons. Because pentagons and hexagons in the mask neighborhood all have acute Angle, this design has the advantage that the pixels are in acute angle of a complex shape field, the mask can also find a homogeneous neighborhood consistent with edge shape basically. The edge details are retained while removing the noise.The research works above laid a theoretical foundation for the engineering practice of the up-down material robot serving for numerical control machine tool group, and those are of great significance not only to improve the level of automation of manufacturing industry,but also to enhance the applicability of computer vision technology in the industrial field.更多还原