【作者】 王华；

【导师】 王龙山；

【作者基本信息】 吉林大学 ， 机械制造及其自动化， 2009， 博士

【摘要】 本文结合企业技术创新项目和吉林省科技应用基础研究基金项目《基于机器视觉的汽车车架纵梁检测系统的研究》(项目编号:20060534),以汽车车架纵梁为具体应用对象,研究了纵梁装配孔的孔径尺寸、位置尺寸和类型识别的视觉在线检测技术。论文分析了图像拼接方法,针对频域相位相关匹配算法,提出了基于区域重叠的图像拼接改进算法,该拼接算法的计算量只与图像尺寸的大小有关,而与图像之间平移量的大小无关,解决了车架纵梁图像拼接时,由于背景噪声的存在而导致的误匹配问题,且能有效地克服光照变化和镜头几何畸变所带来的影响,具有较强的抗噪声能力。针对拼接后的纵梁全景图像,提出了边界保持的加权平滑融合算法,使拼接后的纵梁图像,在保持边界特征的同时,自然缝合,解决了因图像采用简单的加权平滑融合算法而产生的拼接区域边界模糊的问题。为了提高对拼接后纵梁全景图像的处理速度,提出了基于局部区域的装配孔特征提取算法,对符合圆孔条件的区域进行局部区域标记,对圆的特征的提取,只在有装配孔的区域进行搜索,减少了搜索区域,可大大节省装配孔边缘检测时间,提高了纵梁检测速度。提出了一种基于非线性畸变校正和双线性变换相结合的方法,对摄像机内部参数的标定和图像的几何变形进行校正,解决了由于摄像机的成像面和待检测的纵梁平面不平行,有一定倾斜角,使得纵梁上的圆形孔图像变成椭圆孔图像的问题。通过利用标准试件作为参照物进行相对标定,使标定精度、可靠性均能达到设计要求。提出了采用ART神经网络的方法完成多种型号车架纵梁类型的识别,并对ART2神经网络结构和连接权值的算法做了改进,使神经网络能够识别纵梁检测中呈比例关系的输入向量,使网络的识别能力更强。通过将ART2神经网络与D-S证据理论相结合,完成多特征数据的证据积累,有效地降低纵梁识别的不确定性,弥补了因单特征模板信息的不足而无法判别的情况,提高了纵梁类型的识别率。根据纵梁在线检测的技术要求,研究开发了车架纵梁计算机视觉检测系统,对5种型号的纵梁试件进行了装配孔的孔径尺寸、孔的位置和纵梁类型进行了在线检测,检测结果准确。实验表明,本文研究开发的车架纵梁在线检测系统具有较高的检测精度,同时能够满足车架纵梁在线检测速度要求。本文研究所取得的成果,对于应用计算机视觉技术完成汽车车架纵梁质量在线检测具有理论意义和实用价值。更多还原

【Abstract】 Computer vision is a modern detecting technology, the actual engineering project as research its object, computer vision method as its base, and adapts image processing, precise measurement, intelligent control and pattern recognition. It has the characters of gaining a great deal of information quickly, and being apt to integrate with design and manufacture control information. Intelligence, digitization, miniaturization, networking and multi-function can be realized. Hold the ability of online detecting, real-time analysis and control. At present, computer vision detecting technology is used as robot, industrial detecting, object recognition, medical image analysis, military navigation and traffic control etc.Automotive industry is used as the mainstay industry of national economy. Its manufacturing engineering reflects a level of a national basic industry. Whether detecting method and technology of automotive productive enterprise are advanced or not, they have become a very important factor of regulating realization of automobile manufacture and improvement of production quality.The automobile frame is a very important part of automobile steering system. It’s the major bearing structure; at the same time, motor, turn system, transmission system and brake system connect with the auto rack girders through the assembly holes in it. Before the total assembly of automobile, dimension, position and numbers of assembly holes are required to be detected. If auto rack girder with disqualified holes is put on assembly lines, this will lead to some parts or the whole automobile can’t be assembled and so on serious problems. So the detecting results of dimension, position and numbers of assembly holes affect the execution of automobile assembly technology. At present, there isn’t online detecting system of assembly holes quality for auto rack girders in my country. Due to missing manufacture of holes or disqualified dimension and etc, corporations break off production usually. This led to the huge economic loss.The goal of this item is to change the out-dated status of online detecting on domestic auto rack girders quality, and hopes to develop domestic technology and shorten overseas and domestic gap.This paper associated the technology innovative item of corporation with Jilin province technology application basic research funding item“Research of The Auto Rack Girders Detecting System”(item No.20060534), the auto rack girders being used as the applied objects, and researched online detecting technology of dimension, position and kinds recognition of the assembly holes.This paper analyzed the image mosaic method; for phase related match algorithm in frequency domain, introduced an improved image matching algorithm based on overlapped region. Counting quantity is related to dimension of image, and unrelated to translation quantity among images. When carrying on image mosaic, this algorithm solved mismatching because of existence of background noises, conquered the effect of light transformation and lens geometric distortion, and had the stronger anti-noise ability.For having been matching auto rack girder panorama images, this paper introduced a weight smooth fusion algorithm based on boundary maintenance. This algorithm not only can keep boundary character, but also make region boundary clear. To improve the processing velocity, introduced an extraction algorithm of assembly holes based on local region, namely, carry on local region marking for the region of meeting holes condition; for character extraction of circles, only search in the region with assembly holes. By way of this method, decrease the searched regions, save the time of edge detection and improve the detecting velocity of auto rack girders.This paper proposed a method of combination non-linear distortion correction and double-linear transformation, to calibrate camera internal parameters and correct geometric distortion. The method solved the problem of circular holes becoming oval in the auto rack girders, because imaging plane of camera isn’t parallel to the under-detecting auto rack girder plane. Using standard template as reference, through the relative calibration, accuracy and reliability of calibration meet the design demands.This paper proposed a method of ART neural network to recognize the kinds of multi-type auto rack girders, and improved the network structure and joint weight of ART2 neural network. By this way, neural network can recognize the input vectors in proportional relation, recognition of network becomes stronger. Through the combination ART2 and D-S evidence theory, complete the evidence accumulation of multi-character data, to decrease the uncertainty of auto rack girders recognition, make up for the shortage of single-character template information and improve the recognition rate.According to the technology demands of the auto rack girders online detecting, this paper researched and developed the computer vision detecting system of auto rack girders, and found the standard database of auto rack girders images. Utilizing the research results of this paper, on-line detected the dimension and position of assembly holes and kinds of auto rack girders for 5 types auto rack girders test pieces. Detecting results are actual.Experimental results indicate, online detecting system of auto rack girders which was researched and developed by this paper has higher detecting accuracy, at the same time may meet the demands of online detecting velocity. The research outcome of this paper, for online detecting of the auto rack girder with computer vision, has the theoretical and practical value.更多还原