【作者】 王立中；

【导师】 麻硕士；

【作者基本信息】 内蒙古农业大学 ， 农业电气化与自动化， 2009， 博士

【摘要】 奶牛的现代育种技术主要从奶牛的性能和体型两方面指标进行评定,对奶牛体型指标重要性的认识正在逐渐提高。将机器视觉技术应用于奶牛的体型评价体系,不仅可以满足奶牛养殖业的实际生产需求,而且通过机器视觉技术重建获得的虚拟牛可为后续进行的虚拟解剖等研究奠定坚实的基础,新的和改进的算法与模型的提出,丰富了机器视觉理论的研究内容,也为相关领域的应用提供了借鉴和参考。本文从技术路线角度分析了国内外机器视觉技术理论方法及在奶牛体型评定中关键技术的应用。阐述了射影几何原理、相机成像模型、图像预处理、特征点检测技术、特征点匹配技术、相机标定技术等基本理论方法。本文提出一种基于图像分块的多尺度Harris特征点检测算法。以提取图像一定数量的特征点能最大程度地表征图像特征为目的,研究了Harris特征点检测算法。由于Harris算法在特征点检测时阈值是经验性的,对检测要求的特征点数量只能多次尝试,而且检测中阈值是单一的,造成特征点在局部区域聚簇,分布不均匀,不能很好的表征图像整体特征。本文算法在保持原算法计算简便性和稳定性的前提下,对图像分块检测特征点,每块内自动选取阈值。得到的特征点区域可控、数量可控,整体分布均匀,有效地减轻了数据冗余和聚簇现象。本文提出一种改进的SIFT特征点匹配算法。以提高图像特征点匹配算法效率为目的,研究了SIFT特征点描述子基于欧氏最小距离测度的匹配算法。由于SIFT特征点检测算法检测到的特征点数量较大,且每个特征点描述子都是128维的向量,而基于欧氏最小距离测度的匹配算法要求,待匹配第一幅图像的每个特征点要和待匹配第二幅图像的所有特征点求距离,排序后寻找极值,这导致了算法效率较低。本文依据光学成像理论和双目视觉理论,由第一幅图像每个特征点的坐标,从行列两个方向缩小第二幅图像待匹配特征点坐标的搜索范围,本文算法在保持匹配精度的基础上,提高了算法的效率,算法速度约是原算法速度的2.7倍。本文提出一种基于传统相机标定方式的自动标定方法。以提高相机标定效率和自动化程度为目的,研究了相机标定方法。由于传统相机标定需要手工标识,工作量大、效率低。本文相机标定方法使用自制标定模板,应用特征点检测技术代替手工标识,实现自动检测标识点并自动排序,标定精度和原算法基本一致,但可靠性好,实用性较强,简化了相机标定过程。综上所述,本文将特征点检测、特征点匹配和相机标定等三维重建关键技术应用于奶牛体型评定中,旨在得到匹配特征点的三维坐标数据,进而得到奶牛体型参数。更多还原

【Abstract】 Modern appraisal technology of dairy cow is mainly processed from two aspects of capability and conformation,and the latter is more important. It can meet the practical requirement using machine vision technology to the conformation appraisal system of dairy cow, and the virtual reconstruction cow builds sound foundation for virtual anatomizing etc. The new and improved algorithm and model on the study are proposed, which widen the machine vision theory fields, and provide the reference for the correlative fields.Overseas and national machine vision theory and its application on the dairy cow conformation appraisal are analyzed in the dissertation. The theory is expatiated about imaging camera model, image pretreatment, detecting feature points, matching feature points, camera calibration etc.A multi-scale Harris algorithm for feature point detecting based on image sub-block is proposed. The Harris algorithm is studied in order to represent the image feature using a certain number feature points extracted from image. The threshold is selected experientially in detecting feature point using Harris algorithm, resulting in repetitious attempts for detecting a certain number feature points. In addition, selecting simplex threshold induces the clustered feature points in part fields and uneven distributing, so the image feature can not be well represented. In the new algorithm, the threshold is selected automatically to overcome the shortcoming of the original algorithm, which can meet the requirement of feature point’s quantity. The feature points distribute proportionally, so the global image feature is commendably presented. Experiments demonstrate the good performance of the proposed algorithm in controlling fields, controlling quantity, and alleviating the data redundancy and feature points clustered.An improved SIFT algorithm for feature points matching is proposed. The matching algorithm of SIFT feature point descriptor based on Euclidean minimum distance is studied in order to improve the algorithm efficiency of image feature points matching. A large number of feature points are detected with SIFT algorithm, and every point descriptor has 128 dimensions, and distances of every feature point in first image and all the feature points in the second image need to be calculated and then sorted to find the extremum, so the algorithm efficiency is low. According to optics imaging theory and the binocular vision theory, the area of feature points matched in the second image can be reduced from row and column directions based on the feature point in the first image. The improved algorithm accelerates the speed of the algorithm 2.7 times with the same precision yielded.An automatic camera calibration method based on conventional camera calibration is proposed. The camera calibration method is studied in order to improve the efficiency and automation grade of camera calibration. The conventional camera calibration needs marking the marker point by handwork with high workload, low efficiency and unwarranted precision. The new method uses self-made calibration template and detects marker point automatically instead of handwork, which can detect and sort the feature points automatically. Experiments demonstrate that this new method has the same calibration precision as original one, but has steady reliability, strong practicability, and predigests the camera calibration course.In conclusion, the 3D reconstruction key technology including the feature points detecting and matching and camera calibration is applied for conformation appraisal of dairy cow in order to obtain the 3D coordinate data and finally obtain the parameters of dairy cow conformation.更多还原