【作者】 刘洋；

【导师】 李玉山；

【作者基本信息】 西安电子科技大学 ， 电路与系统， 2007， 博士

【摘要】 运动目标提取和跟踪是视觉分析、高级行为理解以及运动编码等计算机视觉应用领域中基本而又重要的一步,也是计算机视觉中的难点问题。本文深入研究了运动视觉中目标的精确提取与跟踪算法,包括运动检测、阴影消除、外轮廓提取以及目标的跟踪四个方面。主要研究成果为:1.针对运动检测门限的自适应选取问题,利用二维时空熵准则自适应确定检测门限;推导出一种迭代计算形式,并利用小波变换对二维直方图进行二次抽样降低门限的搜索范围,极大地降低了计算量;在此基础上,推导出可以利用积分和迭代操作快速计算二维熵门限的新形式,进一步降低了二维熵门限的求解复杂度;并将该方法成功应用于帧差和背景减除方式下的运动检测。2.在原有的熵准则基础上引入模糊理论,形成二维模糊熵准则;给出二维模糊熵门限的快速积分和迭代实现形式,将灰度级为N的图像的各种运算操作次数从O ( N 4)降低到小于或等于O ( N 3);与时空熵门限运动检测相比,时空模糊熵运动检测在提取运动信息和抑制噪声方面有了进一步提高。3.提出一种利用色度畸变和纹理特征进行阴影消除的方法。该方法分析了场景点在存在阴影前后色度的分布规律以及纹理的互相关性,根据亮度信息和饱和度信息选择不同的阴影消除机制,利用颜色向量夹角和纹理特征进行阴影的消除。4.提出了一种彩色图像序列中多运动目标精确外轮廓自动提取算法。采用指数滤波计算彩色图像的梯度强度,改善了彩色图像梯度局部极值点的定位精度,所得到的梯度强度和运动边缘位置信息供随后的轮廓迭代中使用,成功地避免了动态轮廓的过收缩。此外,充分利用轮廓的方向信息,又有效地克服了动态轮廓迭代中的外扩展。从而,自动地提取出一个真正紧贴的目标外轮廓。5.提出一种根据场景变化动态建立目标模型的粒子滤波视觉跟踪算法。选择简单且具有互补性的特征描述表示当前图像,并统一采用直方图法对目标和背景的这些特征进行建模;在粒子滤波框架下,根据巴塔恰里亚距离测度评价的各个特征对目标和当前背景的区分程度动态调整每个特征的置信度;并对各个特征似然函数的噪声参数进行在线估计和更新,使其似然函数的度量标准达到统一,以维护目标和背景的区分度,进一步提高粒子滤波跟踪的可靠性。更多还原

【Abstract】 Moving objects extraction and tracking is a basic and important step in computer vision application field, such as visual analysis, high-level behavior understanding and motion coding, and they are also challenges in research field of computer vision. In this dissertation, accurate moving objects extraction and tracking in motion vision are studied in depth, including motion detection, shadow suppression, contour extraction of moving objects, and objects tracking.The major contributions of this dissertation are summarized as follows:1. In order to determine the threshold for motion detection, two-dimensional spatio-temporal entropy principle is used to determine the threshold adaptively. Aiming at reducing the complexity of traditional two-dimensional entropic thresholding method, a fast iterative solution is deduced, and wavelet transformation is used to decrease the search scope of threshold candidates. On the basis of above research, another new solution for two-dimensional entropy threshold is deduced, in which repeated calculation is avoided by using integral and iterative operation, thus further reducing the computational complexity. This method is applied successfully to motion detection which is based on frame difference or background subtraction.2. Fuzzy theory is introduced to original entropy principle, and fuzzy entropy principle is formed. A fast solution for calculating the two-dimensional fuzzy entropy threshold is deduced, in which the various calculation operation is reduced from O ( N 4) to less than O ( N 3) for N gray-level image. Compared to the two-dimensional entropic thresholding method, this method has better performance on both motion information extraction and noise suppression.3. A shadow suppression method which combines chrominance distortion and texture feature is proposed. The distribution rule of the chrominance and the cross correlation of texture after and before shadow appears are analyzed, different shadow suppression mechanism is selected according to brightness and saturation information, and chrominance distortion and texture feature are used to eliminate shadow.4. An algorithm to extract accurate contours of multiple moving objects in color image sequences is proposed. The exponential filter is used to calculate the gradient magnitude of color image and the localization accuracy of local maximum of gradient in color image is increased, the gradient magnitude and the motion edges are used更多还原