【作者】 李智慧；

【导师】 黄凤岗；

【作者基本信息】 哈尔滨工程大学 ， 计算机应用技术， 2008， 博士

【摘要】 运动目标分割与跟踪在军事、气象、地质、商业传媒等很多领域都有着广泛的应用,是模式识别、图像处理、计算机视觉等领域的重要课题。本文首先对运动目标分割与跟踪的研究现状进行了分析与总结,介绍了该领域现有的研究方法,取得的成就以及存在的问题。本文采用最大后验边缘概率(MPM-MAP)算法实现运动目标分割,MPM-MAP算法是在最大后验概率(MAP)图像分割算法的基础上发展而来的。这类算法既能估计运动参数,又能分割出运动目标,具有结果准确,适应广泛的特点。MPM-MAP算法在形式上明确了通过两步估计运动参数和运动目标所对应的区域(支持区);采用二值标记场表示支持区,从而使计算简化。该算法比MAP算法和期望最大化(EM)算法更灵活,速度更快。本文从MAP运动分割算法入手,详细介绍了MPM-MAP算法的理论框架,还分析了MPM-MAP算法与MAP算法及EM算法的区别。在此基础上本文对该框架中的数据平滑算法——基于马尔可夫随机场(MRF)模型的算法进行了改进。改进后的算法利用MPM-MAP算法中的二值标记表示法,统一更新图像标记场数据,通过均值滤波器实现MRF能量函数的计算,在保持算法平滑效果的同时提高了MRF算法的运行速度。传统的MPM-MAP算法虽然运行速度比相近的MAP算法和EM算法有较大提高,但仍存在一些缺点。如缺乏有效的运动目标数和初始运动参数的估计方法;常采用的估计运动参数的最优化方法依赖于初值,对目标函数要求可导等。MRF平滑方法虽然效果很好,但用于迭代过程中时,即使是快速算法仍然运行时间较长。针对以上缺点,本文提出了以下改进方法。本文改变了MPM-MAP算法中数据平滑项的定义,定义平滑项为运动目标备选像素的密度,选择备选像素密度最大的矩形区域为运动支持区。采用区域收缩算法实现图像平滑、去除噪声,该算法不仅可以消除噪声,还可以对运动目标定位并计算运动目标的特征矩形。区域收缩算法相对MRF算法具有运行速度快的优势。除可以用于支持区估计结果上,区域收缩算法也可以用于差分二值图像,本文采用连通区域包围盒与区域收缩算法相结合的方法估计初始支持区,将传统算法先假设运动参数再估计运动支持区的顺序改为先估计支持区后估计运动参数,提高了初始参数估计的准确性。本文选择运动模型为6参数仿射模型。在区域收缩算法的基础上,本文提出一种基于矩形区域主轴的轴仿射模型。该模型通过矩形区域的两个主轴估计仿射运动参数,其优点主要有以下几点。一是在不影响运动参数估计准确性的前提下提高了运动估计的速度;二是使仿射运动参数具有明确的几何意义,容易确定一个初始参数的范围,使参数估计的最优化算法除传统的基于梯度的方法及随机最优化方法外,可以使用搜索类的算法。本文选择一种较新的取值有界的最优化算法—DIRECT算法实现运动参数的估计。该算法不要求目标函数可导,也不用假设初值。所以,轴仿射运动模型与DIRECT算法的结合提高了参数估计的准确性、稳定性和鲁棒性。本文对上述算法与模型进行了详细介绍。最后,本文将上述算法应用于运动目标跟踪。本文的跟踪算法采用矩形表示法,基于区域收缩的MPM-MAP算法用于估计初始运动目标及运动参数。目标跟踪采用基于卡尔曼滤波的状态空间目标跟踪方法。状态为轴仿射参数,DIRECT算法的中间结果用于计算观测向量的协方差矩阵。实验结果表明,矩形表示法、轴仿射模型和DIRECT算法的结合有利于更好地实现运动目标跟踪。本文提出的算法都经仿真实现,并证明可行。文中给出较详细的实验结果。更多还原

【Abstract】 The segmentation and tracking of moving object are widely applied in many fields such as military affairs, weather, geography, commerce and media. They are very important research topics in pattern recognition, image processing and computer vision. The research status of moving object segmentation and tracking is analyzed and summarized in the dissertation. The approaches, achievements and problems of current works in this field are introduced then.The maximizer of the posterior marginals-maximum a posteriori (MPM-MAP) algorithm is adopted to implement moving object segmentation in the dissertation. MPM-MAP is developed on the basis of maximum posterior probability (MAP) algorithm. This kind of algorithms can both estimate motion parameters and segment moving objects. These algorithms are featured by their accurate results and wide applications. MPM-MAP algorithm makes it clear in the form that motion parameters and regions corresponding to the moving objects (supporting regions) are estimated in two steps. The binary labeling field is used to represent the supporting regions, which simplifies the calculation. The MPM-MAP algorithm is more flexible and faster than MAP and expectation maximization (EM) algorithms.With the MAP algorithm as the starting point, the theoretical frame of MPM-MAP algorithm is introduced in detail in the dissertation. The differences in MPM-MAP algorithm, MAP algorithm and EM algorithm are also analyzed in the dissertation. The smoothing algorithm in MPM-MAP based on Markov random (MRF) model is improved in the dissertation. In the improved MPM-MAP algorithm, the labeling field data are updated uniformly using the binary labeling representation and the MRF energy is computed by mean filter. Therefore, the improved algorithm heightens the speed of MRF algorithm while the smoothing effects remain the same as the traditional algorithms.Although the traditional MPM-MAP algorithm is much faster than MAP and EM algorithm, some shortcomings still remain in it. For example, it lacks of the effective approaches to estimate the number of moving objects and the initial motion parameters; the most often used optimal methods for motion parameter estimation are dependent on the initial values; the objective functions are required to be differentiable. MRP model has a good smoothing effect but it seems to be slow in running speed during the iterations even though for the fastest MRF algorithm.Aiming at the shortcomings mentioned above, some improvements are proposed in the dissertation. The smoothing term in MPM-MAP algorithm is changed and defined to be the density of pre-selected pixels belonging to the moving objects. The rectangular region with the maximum density is chosen as motion supporting region. The region shrinking algorithm is used to implement image smoothing and noise suppression. In addition to the noise suppression, this algorithm can locate the moving objects and compute the feature rectangles of the objects.The region shrinking algorithm has the advantage of higher speed than MRF algorithm. It can also be used in binary difference images as well as the estimation of supporting regions. The bounding boxes of connected regions are combined with region shrinking algorithm to estimate initial supporting regions. The original algorithm calculates the motion parameters first and then supporting regions. Supporting regions are computed first and then the motion parameters in the proposed algorithm. Therefore the initial parameters are more accurate than the traditional algorithm.The 6-parameter affine model is chosen as motion model in the dissertation. An axial affine model based on rectangular region is presented. This model describes affine motion by center translation, rotation and scaling of two main axes of a rectangular region. Its advantages are as follows. The first advantage is heightening the speed of motion estimation while keeping the accurate results as before. The second advantage is that it makes a clear geometric meaning for the affine parameters. The range of initial parameters is easy to be obtained. The optimal methods of parameter estimation can be chosen in searching approaches except the gradient-based and random optimal methods used before. A comparatively newer optimal approach with limited parameters—DIRECT algorithm is used to calculate the motion parameters. The method does not need a differentiable objective function and a set of assumed initial parameters. The combination of axis affine model and DIRECT algorithm improves the precision, stability and robustness of parameter estimation.The model and algorithms mentioned above are described in detail in the dissertation. Finally these algorithms are used in moving object tracking. The representation of rectangular region is adopted in tracking. The initial regions of objects and motion parameters are computed by the MPM-MAP algorithm based on region shrinking. The tracking approach based on kalman filter is presented in the dissertation. The state of kalman filter is a set of axial affine parameters. The covariance matrix of the observed vectors is computed by the interim results of DIRECT algorithm. The experiment results show that the combination of rectangular representation, axial affine model and DIRECT algorithm favors the better tracking of moving objects.The algorithms in the dissertation are simulated and proved to be feasible. The detailed experiment results are also given in the dissertation.更多还原