【作者】 田鹏辉；

【导师】 隋立春；

【作者基本信息】 长安大学 ， 资源与环境遥感， 2013， 博士

【摘要】 论文主要关注视频图像中运动目标检测和运动目标跟踪相关理论方法的研究，即如何让计算机从视频图像序列中获得物体运动数据。利用不同的数学模型工具建立视频图像中单纯的理想背景图像，一直以来，是运动目标检测领域所追求的最终目的。针对实际视频图像中复杂的背景环境（光照变化、有微动物体等），在分析传统背景建模方法的特点与不足后，提出一种基于Kalman滤波思想的自适应背景建模运动目标检测方法，相对于平均背景建模和混合高斯建模，本文方法能够适应比较复杂的背景环境，提取出的背景图像接近于真实的背景场景。以此运动目标检测方法为依据，同时开展了有关运动目标跟踪领域的研究，针对快速视频图像中前后帧图像中运动目标没有重叠或者重叠区域较小问题，粒子滤波目标跟踪中粒子数量选取和粒子发生退化现象的问题以及运动目标存在大小和姿态变化的目标跟踪问题，提出了三种不同的运动目标跟踪方法，同时通过仿真实验对比，验证了每种方法的适应条件和实际的跟踪效果。论文借助高斯模型理论、kalman滤波思想、均值漂移算法、贝叶斯理论、小波函数理论、粒子滤波方法和傅里叶-梅林变换（Fourier-Mellin Transform，FMT）等方法，对视频图像中运动目标检测与目标跟踪中存在的关键问题开展了一些研究工作。本文的主要研究内容和结果如下：（1）在前景区域检测与提取中，讨论各种常用的背景建模方法原理时，重点分析了混合高斯模型背景（Mixture of Gaussian，MoG）建模方法，提出了基于kalman滤波的背景建模运动目标检测算法。该方法是一种递归的思想，利用前一帧图像和当前观测图像计算出当前帧背景的估计值，具有无偏性、迭代性以及稳定等优点。算法克服了传统的平均背景模型（Average Background Model）建模方法对场景中的光照变化和背景的多模态性比较敏感的不足，背景建模的计算量又小于基于统计理论的MoG建模方法。（2）本文对传统的用于目标跟踪的均值漂移（Mean Shift，MS）算法进行了研究和分析，在此理论的基础上，提出自适应均值漂移目标跟踪算法。相对于传统MS算法采用Bhattacharyya系数来度量模板图像和观测图像的相关性，本文算法在二维概率分布图像中，计算目标图像区域内的0阶矩和两个1阶矩，从而不断计算目标区域的质心，并调整位置，使得质心达到运动目标的真实中心。该方法克服了传统MS算法中前后两帧图像中目标没有重叠或者重叠区域较小时跟踪失效的不足。（3）对经典粒子滤波目标跟踪算法进行了细致分析研究，针对粒子滤波目标跟踪中粒子数量选取多少和粒子发生退化现象的两难问题，提出融合颜色和纹理特征的粒子滤波目标跟踪算法。相对于传统的单一特征粒子滤波跟踪算法，该方法只需要较少的粒子数就可以达到理想的跟踪效果，同时对粒子退化现象也有一定的改进。（4）在研究经典图像匹配算法的基础上，结合可见光图像处理的自身特点，提出基于傅里叶-梅林变换的运动目标跟踪算法。该方法基于图像匹配的思想，把待跟踪目标与实际的观测图像进行图像匹配，从而估计出跟踪窗口的大小和位置，同时利用图像匹配的平移、旋转和缩放参数更新待跟踪目标模板，实现了运动目标大小和姿态变化的目标跟踪。更多还原

【Abstract】 This thesis is concerned with moving single target detection and visual trackingalgorithms and their application in the real world. For a long time, it is the ultimate goal toestablish ideal background image by different mathematical mode tools in moving objectdetection field. According to the actual complex background environment (Lighting change,subtle moving, etc.) in video images, an adaptive background modeling method is proposedfor moving object detection based on Kalman filter theories, this method can be adapted tocomplex background environment, extracting background image close to real backgroundscene. At the same time, three different methods of moving object tracking is presented, thoseis an adaptive Mean Shift tracking algorithm, a tracking algorithm fusion of color andtexture features under the framework of particle filter and the algorithm of moving objecttracking based on Fourier-Mellin Transform. The experiments result is that each method isrobust and adapted to the different conditions.With the help of the Gaussian model theory, kalman filtering ideas, mean shift algorithm,Bayesian theory, wavelet function theory, particle filtering method and Fourier-MellinTransform(FMT) method, this paper carry out a lot of research work in moving objectdetection and tracking of the video images or image sequence.In this paper the main researchwork and innovative ideas are as follows:(1) In part of detecting and retrieving the foreground, this paper analysis several genericbackground modeling methods， especially for the mixture Gaussian model in detail, analgorithm of moving object detection is proposed based on kalman filtering backgroundmodeling. The method estimates the background of video images or image sequence by theprevious frame image and the current observation image, and it is iterative, stable andunbiased. The presented method overcomes shortcomings of the traditional averagebackground model (ABM) method that it is sensitive to light changes and multi-mode distribution of background pixel. The amount of calculation is less MoG method based onstatistical theory.(2) Researching on the theory of Mean Shift (MS), an adaptive tracking algorithm is proposedbased on mean shift. In the traditional MS tracking method, the correlation between templateimage and observed image is measured by Bhattacharyya coefficient. In this paper, the objectcentroid is over and over again calculated by0-order moment and tow1-order moment whichgot from two dimensional gray histogram of target images area. The position of the trackedobjects is adjusted based on the object centroid so that it moves to the real target center. Thismethod overcomes the deficiency of the traditional MS tracking algorithm when the overlapregion of tracked object in the adjacent frames is too small or no overlap area to track thetarget.(3) Researching on objects tracking algorithm based on particle filter, the tracking algorithmis proposed to fusion of color and texture features under the framework of particle filter. Thepresented method requires less number of particles can achieve the desired trackingperformance compared with the single feature tracking algorithm in particle filter, and itimproves the particle degeneracy in certain extent.(4) Researching on the images matching algorithms and the characteristics of visible lightimages, the algorithm of moving object tracking is presented based on Fourier-MellinTransform. The method is based on the idea of image matching, it can estimate the size andposition of the tracking window by matching the tracking object with the observed image, andupdate the tracking target template by the translation, rotation and zoom parameters based onimages matching, the proposed algorithm achieved tracking moving object which size andposture changed in moving.更多还原