【作者】 李波；

【导师】 袁保宗；

【作者基本信息】 北京交通大学 ， 信号与信息处理， 2012， 博士

【摘要】 基于视频的运动目标检测与跟踪分析一直是计算机视觉领域中最活跃的研究方向之一,它是人机交互、智能监控、机器视觉导航、工业机器人等应用的基础和关键技术。运动目标检测跟踪是利用图像序列中运动目标在颜色、边缘、纹理、时空等方面上的差异,检测出运动物体；提取目标轮廓形状等信息,并获取目标的位置、尺寸、速度等运动状态,进行目标跟踪。然而,由于现实环境中的阴影、光照变化、复杂背景、目标变化、遮挡等影响,使得进行精确并且鲁棒的运动目标检测跟踪非常困难。本文对视频序列中运动目标检测和跟踪技术进行了研究和探讨,针对第四代人机交互中运动目标检测跟踪遇到困难提出了一些新的算法,主要内容如下：(1)针对阴影对运动目标检测的影响,提出了一种用阴影流和3D马尔可夫随机场后验概率最大化方法(3D MAP-MRF)消除运动阴影的方法,将运动检测和阴影消除一起处理,可在检测结果的同时直接消除阴影。1)首先对场景中每个像素建立混合高斯模型,通过一个阴影弱分类器,比较混合高斯模型中的背景高斯和当前帧图像,将可疑的阴影像素分离出来,送到阴影流模型中。在线学习阴影弱分类器分离的可疑的备选阴影像素,建立阴影流模型。2)用混合高斯模型,阴影流模型和前后帧图像一起构建一个三维的马尔可夫随机场模型,并提出一种新的能量函数形式。通过证明构造的能量函数满足F2定理,是“图可表示的”,可以通过图切法得到最优解。3)构建一个与3D MRF对应的3D网络流图,通过动态图切算法,求出图的最小切,即求得马尔可夫随机场的标号组的最大后验概率。从而给每个像素分配一个不同标号达到分割检测运动物体的目的。实验表明算法可以很好地消除运动阴影的同时,得到较精确的运动检测结果。(2)针对复杂的动态场景,我们提出了基于累计局部核直方图非参数估计的运动目标检测算法。将像素周围历史的纹理信息通过累计局部核直方图的形式表现出来,实验表明可以有效地在水波纹、树叶晃动等复杂动态场景较鲁棒地提取运动目标。该算法首先针对每个像索计算出以其为中心的周围邻域的局部核直方图,然后统计出前N帧对应像素点的累计局部核直方图,以非参数核密度估计的形式,对累计局部核直方图进行非参估计,求得更为精确的非参估计的累计局部核自方图。然后通过与当前帧应像素的局部核直方图进行比较,度量当前像素的局部核直方图和非参估计的累计局部核直方图的距离。通过衡量周围像素局部核直方图的相似性,减少动态纹理,摄像机抖动的影响。该方法可以很好的处理复杂背景中动态纹理的影响,对背景动态变化较大时,也有较好的鲁棒性。(3)提出了一种新的码书和纹理特征结合的运动目标检测方法,通过在线学习构建码书纹理背景模型,解决动态纹理、轻微光照变化等问题。首先用码书以类似聚类的方式构建每个像素的码书模型,根据码字的颜色和亮度相似性,将背景像素分布用聚类码字的形式表示出来,同时在运动检测过程中不断对码字进行重新排列和剔除,更新码字以反映背景变化。然后用单高斯模型来学习背景像素变化的概率,生成高斯局部二值模式(GLBP)纹理算子,同时在线更新GLBP反映图像空间纹理信息变化。最后融合各个特征将当前帧分割为前景背景两部分。实验表明本算法取得了较好的检测效果。(4)针对跟踪目标变化及遮挡问题,提出了一种基于LMSF-MS的改进MeanShift目标跟踪算法,对目标模型进行更新以适应目标变化。用LMS自适应滤波器对核直方图分量进行滤波,在跟踪的同时动态更新目标模型的核函数直方图非零分量,剔除目标模型中已消失的分量,加入目标模型中新出现的分量。使目标模型更能准确地反映目标变化,为模型匹配提供更好的目标模型,跟踪准确,并且速度快。实验表明对于目标有变化、遮挡等情况,跟踪结果准确且鲁棒。本文针对目前运动目标检测跟踪中遇到的阴影、动态复杂背景、光照变化,目标变化及遮挡等问题,提出了相应的解决算法,实验证明我们的算法取得了较好的效果。更多还原

【Abstract】 Moving object detection and tracking is one of the hotest research topic in computer vision field. It is the basic and key technology of human-computer interaction, intelligent control, and visual navigation. Moving object detection is a process of finding motion in image sequences using color, edge, texture, space and other aspects of differences, extracting the shape of object, locating the object’s coordinates, size, speed etc., and tracking it. However, in real environment, it is very difficult to detect the moving object accurately and robustly, because of shadow, illumination change, complex background, objective change and occlusion.Video based motion detection and tracking technology has been studied and discussed in this thesis. For some difficulties in the fourth-generation human-computer interaction, we introduced some new algorithms to resolve them. The main contributions of the dissertation are as follows:(1) To remove moving shadow, we present a novel approach of moving shadow elimination based on Shadow Flow and maximum a posteriori probability of3D Markov Random Field (3D MAP-MRF) by integrating motion detection and shadow elimination.1) Gaussian Mixture Model is built as background model for each pixel. By comparing current pixel and GMM, we classify candidate shadow pixel through a weak shadow classifier and send it to Shadow Flow Model. Which isbuilt through the online learning of candidate shadows coming from weak classifier.2) A3D MRF is constructed of GMM, Shadow Flow and current images. We proved that our new energy function meets F2theorem, so that our model is "graph representable", and can be solved with graph cuts optimally.3)A3D graph is constructed according3D MRF. A dynamic graph cuts algorithm is used to find the min-cut/max-flow, which is equal to the maximum posteriori probability of labels. Each pixel is assigned by "foreground" and "non-foreground" label, and moving object detection with shadow elimination is completed. Experiments show that our algorithm can eliminate the moving shadows and get accurate results.(2) For complex dynamic scenes, we present an approach to segment moving objects with nonparametric estimated cumulative local kernel histogram (NPE-CLKH). Texture information of surrounding pixels is integrated into cumulative local kernel histogram. Experiments show that our algorithm can detect motion robustly in complex dynamic scenes containing the rippling water, leaves shaking etc..By using the correlation and texture of spatially proximal pixels, a local kernel histogram background model is constructed. Cumulative local kernel histogram of the corresponding pixel of N frames is computed. Then probability distribution of cumulative local kernel histogram is estimated with nonparametric techniques. We employ Bhattacharyya distance to measure the similarity of local kernel histogram between estimated background model pixel by pixel in current frame, and compare CLKH of Neighboring pixel, to reduce the impact of dynamic texture, and camera shaking. Our approach can reduce false detections due to disturbing noise and small motions in dynamic scenes. This approach can deal with complex dynamic texture problem, and get robust results.(3) We present a novel texture-based algorithm to detect motion with codebook and Gaussian local binary patterns (GLBP), which can get texture background model on-line, and resolve the problems of complex background, slight illumination changes etc..Firstly, a codebook model is constructed in similar manner of pixel clustering. Distribution of background pixels is represented code words cluster according to using the color and brightness similarity between codebook and current pixel. Our algorithm updates the codebook model both in initial step and detection step to deal with changes of background pixels. A single Gaussian model of pixel-wise is used to build the pixel’s value change model on-line. A background model based on Gaussian local binary patterns is constructed on-line by applying the correlation and texture of spatially proximal pixels. Finally current image is segmented into two parts, foreground and background by fusing those features. Experiments show that our algorithms achieve good detection results.(4) To deal with the issue of target change and occlusion, we propose a novel adaptive model update method for real-time mean shift blob tracking. We use adaptive LMS filter for filtering object kernel histogram and update the vectors of the object kernel histogram dynamically. New bins of target kernel histogram are added, and useless bins are removed. The target kernel histogram can reflect the target changes, and better matchs the target model. Experiments show that our method gets accurate and robust tracking results in the case of target change and occlusion.In this thesis, corresponding solutions are proposed to reslove the problems of shadow, dynamic complex scenes, target change and occlusion. Experiments demonstrate that our algorithms achieve good results.更多还原