【作者】 王选贺；

【导师】 刘济林； 于慧敏；

【作者基本信息】 浙江大学 ， 通信与信息系统， 2011， 博士

【摘要】 长期以来,人体目标跟踪是计算机视觉研究的热点。它是使用计算机及相关设备对生物视觉的一种模拟,是一项非常具有挑战性的研究工作。人体目标跟踪技术目前仍然处于研究和探索阶段,在理论和实际应用中仍存在许多问题没有得到很好地解决。影响目标跟踪算法鲁棒性的原因很大程度上是由于目标运动的复杂性以及光照条件变化等因素造成的。本文主要研究了单摄像机和多摄像机下的人体目标跟踪,单摄像机人体目标跟踪是跨摄像机人体目标跟踪的基础,只有单摄像机人体目标跟踪正确的前提下,跨摄像机跟踪才会有正确的跟踪结果。在单摄像机人体目标跟踪过程中,主要解决遮挡人体目标跟踪。跨摄像机人体目标跟踪分为有公共视场人体目标跟踪和无公共视场人体目标跟踪。本文的主要创新包括：1.针对多人体目标跟踪而产生的遮挡问题,提出了一种基于非参数复合模型的粒子滤波方法来描述多个人体目标的情况。该复合模型通过复合预测和复合更新两个步骤交替进行以达到递归跟踪的目的。在复合粒子滤波过程中,通过每个复合粒子的权值大小来衡量粒子的贡献大小。权值的大小是基于HSV直方图的多颜色观测模型来实现的。观测模型是通过巴氏距离的核密度估计来建模。实验结果证明复合模型的粒子滤波方法可以很好地实现多人体目标跟踪。2.充分利用单摄像机的人体目标的运动信息、颜色信息、空间信息,将背景建模、块建模、颜色建模、运动建模和人体的空间信息进行有效地融合,较好地解决了人体目标相互遮挡情况下的人体跟踪的问题。本文利用人体目标的运动信息,采用的混合高斯模型进行背景重建的方法,先提取出运动的目标；利用基于Epanechnikov核密度梯度估计算法对存储模型中的人体进行聚类,即对人体的外部颜色相近的像素进行块建模；利用非参数的Gauss核密度估计算法对已聚类的块模型和人体的空间信息建立颜色密度函数,同时利用人体运动信息建立运动密度函数,通过颜色密度函数与运动密度函数构成后验概率模型；然后对当前的检测出的人体目标的每个像素计算最大后验概率,得到最大后验概率的颜色图像,通过该图像进行分割,达到对遮挡情况下多人体目标进行跟踪的目的。3针对有公共视场跨摄像机人体目标跟踪误匹配的问题,提出了一种基于空间映射变换与颜色特征信息相融合的匹配算法。本文先通过跨摄像机的公共视场分界线来初步确定跨摄像机的人体目标,然后利用变换单应矩阵计算跨摄像机人体目标的对应关系,然而在实际应用中,视频是不完全同步且有交叉遮挡而产生误匹配,因此本文在投影变换单应矩阵的基础上采用基于颜色模型的最大后验概率进行匹配。4.本文针对无公共视场跨摄像机人体目标跟踪因颜色差异较大而引起误匹配问题,提出了一种基于颜色转变函数的方法对颜色进行校正,并通过颜色转变函数空间进行概率估计来提高跨摄像机人体目标的匹配的准确率。该算法先在低维度的子空间上对已知的跨摄像机人体目标进行训练,得到颜色转变函数。该方法不用依赖摄像机的内参来计算颜色转变函数的子空间。用基于概率的主要成分分析法对颜色转变函数的子空间进行建模得到概率密度函数,最后利用概率密度函数获取跨摄像机的人体目标的匹配概率。本文在室内外,全部室外环境下对该方法进行了实验,从实验结果上看,本方法能够很好地对室外及室内外环境下的跨摄像机人体目标进行跟踪。更多还原

【Abstract】 Human tracking has been a hot spot in the field of computer vision for a long time. It is to simulate the function of human vision using the computer and relative device.It is an extremely challenging research and comprehensive subject. The technolegy of human tracking is still in the course of research and exploration at present. There are many problems to be sloved in the aspect of theory and practical application. The reasons such as the motion complexity of human, alteration of illumination intensity and others influences the robustness of tracking human to a great extent.The dissertation mainly research on human tracking in a single camera and across multiple cameras. The human tracking across multiple cameras is on the basis of human tracking in a single camera. Only on the accurate premise of tracking human in a single camera, the tracking human across multiple cameras may be correct result. The dissertation mainly solves the tracking human under occlusion in a single camera. There are two cases of tracking human across multi-cameras.One is human tracking across multiple cameras with overlapping field of view.The other is human tracking across multiple cameras with non-overlapping field of view. The two cases will be talked about in the dissertation.1. The dissertation aims at solving the problems of tracking occluded human object. An algorithm is proposed based on particle filter of non-parameter mixture model which is described the multiple human objects in a single camera. The mixture model carries on two steps of mixture predict and mixture update by turn in order to accomplish recursive human tracking.During the course of mixture particle filter, the contribution of each particle to final object is weighed using weight of each particle.Each particle weight gets attained utilizing mutiple observation model of HSV histogram. Observation model gets acquired by kernel estimation of Bhattacharyya distance.The experiment shows that particle filter of mixture model can better track multiple human object under occlusion.2. The dissertatation makes full use of motion information,color information and space information of human object in a single camera.The dissertation effectively solves the problems of human tracking under occlusion each other using background modeling,blob modeling,color modeling,motion modeling and sapce information of human body. The dissertation firstly detect the kinetic persons using motion information to reconstruct background of video based on algorithm of mixture Gauss modeling. We will cluster to human body in storage based on Epanechnikov kernel density gradient estimation.Namely, we will obtain blob models,which put the color-similar pixels of human body together.We will establish the color density function based on non-parameter Gauss kernel density estimation and space information of human body, build motion density function based on motion information of human and obtain posterior probability utilizing color density function and motion density function. We apply maximum posterior probability to each pixel of detected human in current frame,obtain color image of maximum posterior probability and segment the color image to aim at tracking multiple human under occlusion in a single camera.3. The dissertation aims at researching on mis-match problem of tracking human object across cameras with overlapping field of view. A fusion algorithm is proposed based on space map transformation and color feature to accomplish match across cameras.The dissertation firstly ascertains the position of human across camera roughly using the common boundary lines of overlapping field view.Then the person across cameras gets matched by means of homography transformation matrix.However,in practical application,the videos across cameras is asynchronous and occluded persons, so there exist some mis-matches.The dissertation calibrates the mis-match by maximum prosterior probability of color model on the basis of homography matrix.4. The dissertation aims at solving mis-match problem caused by color differences across cameras with non-overlapping field of view. An algorithm is presented to calibrate the color pixels based on color transfer function and apply probability estimation to the space of color transfer function in order to advance match accuracy of tracking human across cameras.The algorithm firstly trains the given samples of human objects across cameras in a low dimension to acquire color transfer function.The algorithm is not dependent on internal parameter of cameras to acquire color transfer function.Then the probability estimation function is modeled utilizing principal compoment analysis to sub-space of color transfer function.Finally,match probability of tracking human across cameras using probability estimation function. The experiments in the case of outdoor-outdoor and indoor-outdoor shows the algorithm can better track human object across camera with non-overlapping field of view.更多还原