【作者】 邹腾跃；

【导师】 唐小琦； 陈吉红；

【作者基本信息】 华中科技大学 ， 机械电子工程， 2013， 博士

【摘要】 机器视觉作为光电技术的一个特定应用领域，已快速增长并发展成为一个前景光明、活力无限的行业。本文在国家科技重大专项的资助下，围绕机器视觉应用中广泛使用并至关重要的目标检测及跟踪算法进行研究，并结合其应用展开讨论，主要研究内容和创新点有：针对已知目标物体特征的情况，提出基于加权颜色直方图的目标检测方法。该方法利用像素的纹理信息对颜色特征进行加权生成二维颜色特征直方图，并用该直方图在图像空间中搜索需要的目标。图像实验证实了该方法能取得优于传统算法的检测效果，本文还在并联机器人表壳工件分拣系统上验证了其有效性。针对工件分类的应用需求，提出了基于加权扩散形状上下文的目标分类方法。该方法从传统形状上下文分形因子对图像扭曲噪声敏感的缺点出发，提出使用曲率加权扩散的方法来弥补这一缺陷，并采用动态规划特征点匹配方法来降低算法的时间消耗。该分类算法在表壳工件的分类识别中取得了良好效果。针对未知目标物体特征的情况，提出基于韦伯特征背景建模的目标检测方法。该方法采用韦伯局部描述因子作为特征信息，对每个像素点进行核密度估计背景建模，并用样本更新机制和自适应方差策略来增强算法的鲁棒性。实验基于准确性和鲁棒性两种评价指标展开，并在红外夜视和流水线皮带传送两种实际应用环境中检验了该算法的有效性。针对视频场景目标跟踪问题，提出了基于稀疏表达的粒子滤波跟踪方法，该方法在粒子滤波跟踪框架内实现了局部稀疏表达的目标建模方法，并采用加速最近梯度法来提高求解的实时性。模板在线更新和稀疏字典更新策略被用于改善算法的鲁棒性，粒子滤波估计过程采用系统重采样方法来克服粒子退化问题。实验证实该算法有较好的跟踪精度和鲁棒性，并能满足皮带传送环境下的跟踪需求。本文还基于提出的目标检测与跟踪算法以及华中数控HNC-08型数控系统设计了一套工业机器人目标跟踪拾取系统，并利用MOTOMAN SK6型工业机器人在皮带传送机上进行了相应的物料拾取实验。更多还原

【Abstract】 Machine vision as a specific field of photoelectric technology has rapidly grown to bea promising industry. On the support of the National Science and Technology MajorProject, this paper studies the most important and widely used algorithms with itsapplications in the object detection and tracking field. The major research and innovation:For the case of a target with known features, the paper proposed an object detectionmethod based on the weighted color histogram. By the texture information of pixels, thecolor feature can be weighted to generate the two-dimensional color feature histogram.Then, the detection method uses it to search the desired object in the image space. Theexperimental results confirmed that this method can obtain better performance than theconventional algorithms. This paper also verified its effectiveness on a robot workpiecesorting system.For the application requirement of material classification, this paper proposed aclassification algorithm based on the weighted diffusion shape context which uses thecurvature weighted diffusion mechanism to compensate the effect of image distorted noise.Moreover, it uses the dynamic programming for feature point matching to reduce thecalculation time cost. The algorithm achieved good results in classification of workpieces.For the case of a target with unknown features, the paper proposed a backgroundmodeling method based on Weber Local Descriptor and the Kernel Density Estimation.This method makes kernel density estimation for each pixel using Weber Local Descriptoras its feature information, and then designs the sample update mechanism and adaptivevariance to enhance the algorithm robustness. The experiments have been carried oninvolving the accuracy and robustness of the algorithm. The infrared night vision and beltconveyor application were also tested.For the object tracking problem in video scene, the paper proposed a particle filtertracking algorithm based on adaptive sparse expression. This algorithm constructs thelocal sparse express modeling within the particle filter tracking framework, and uses theaccelerated gradient method to improve the time consumption. The online templatesupdate and sparse dictionary update strategy were proposed to improve the robustness ofthe algorithm. The system re-sampling mechanism is used to overcome the problem ofparticle degeneracy. Finally, the experimental results confirmed that the algorithm had better tracking accuracy and robustness, and could meet the tracking requirements underbelt conveyor environment.An industrial robot target tracking pickup system was designed by the object detectionand tracking algorithm in this paper based on a Huazhong CNC HNC-08system. Animplementation of this system on a MOTOMAN SK6industrial robot was built for pickupexperiments on belt conveyors.更多还原