【作者】 李峥；

【导师】 桑农；

【作者基本信息】 华中科技大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 近年来,智能视频监控技术引起了越来越多研究人员的重视,然而该技术的发展却遇到了多方面的制约,遮挡目标跟踪就是其中之一。在单摄像机的视频监控中,由于观察角度等原因,目标之间的相互遮挡是普遍的现象。这对目标跟踪算法准确跟踪目标带来很大影响,甚至严重影响视频监控系统的应用。因此,本文重点研究遮挡目标的跟踪问题。本文利用贝叶斯理论为目标跟踪问题建模,并对该模型进行理论推导,分别为四种不同的跟踪类型(目标进入场景、离开场景、单目标跟踪和遮挡目标跟踪)推导出表达式。在多目标跟踪中,由于目标自身的情况和目标之间情况的不同,目标跟踪的难度也存在很大差异。为了度量目标的跟踪难度,本文提出不可跟踪性理论,描述三种不同的定义形式:整个图象序列所有目标的不可跟踪性、相邻帧之间所有目标的不可跟踪性以及相邻帧中单个目标的不可跟踪性,并提出了不可跟踪性的简化计算方法,通过实验验证了影响不可跟踪性的四种因素(场景中目标数量、目标的分辨率、目标的运动速度和跟踪特征之间的可区分性)。利用不可跟踪性理论作为指导得到了两种解决遮挡目标跟踪问题的方法:遮挡目标的自动组合和遮挡目标跟踪中的动态特征选择。由于前面提到的跟踪遮挡目标的方法具有一定的局限性,因此本文提出了一个通用的解决遮挡目标跟踪问题的方法。该方法利用遮挡分层的概念来描述目标之间的遮挡关系;利用目标重叠区域的外观特征和速度特征来帮助确定目标遮挡关系;利用目标的遮挡关系和状态信息来获取遮挡目标的非遮挡部分;利用目标非遮挡部分的外观特征和速度特征的组合来描述遮挡目标;对传统的均值平移跟踪算法进行改进,利用外观直方图和速度直方图信息来确定目标的位置;并利用目标在当前时刻以前的目标尺度变化来预测当前时刻目标尺度变化的各个可能性的概率,并通过对尺度变化的概率进行采样来确定目标的尺度变化。描述目标状态的参数包括目标的位置、尺度和目标之间的遮挡关系,这些参数之间互相关联,而且目标的参数空间包括离散变量和连续变量,参数空间随目标数量的变化而变化,因此本文应用马尔可夫链蒙特卡罗方法来求解相关遮挡目标的最优状态。在求解最优状态的过程中,本文对目标模型中的位置、遮挡关系和尺度三个方面的参数构建状态转移函数以加速算法收敛。在算法收敛过程中,三个方面的参数通过采样被逐步地调整以消除这些参数之间的关联,保证算法收敛到全局最优。当跟踪算法收敛时,则认为此时模型的状态参数为目标的最优状态。本文利用多段具有遮挡目标的图象序列对提出的跟踪遮挡目标的方法进行测试,从实验结果来看,本文的跟踪算法能够对所选取的大多数情况的遮挡目标进行很好地跟踪。更多还原

【Abstract】 In recent years, the technology of intelligence video surveillance is attached more importance to researches in computer vision. But in this technology there are many difficulties which include occlusion target tracking. In video surveillance which uses single camera, target occlusion is a common phenomenon due to camera views. The occlusion among targets causes great disturbances for accurately tracking occlusion targets and even for the application of video surveillance.In this paper, Bayesian theory is used for modeling target tracking and the tracking model is represented as different tracking types (target entrance, target exit, single target tracking and occlusion target tracking).In multiple target tracking, the tracking difficulties are difference according to different target situations. In order to measure tracking difficulties, intrackability theory is proposed and three kinds of concepts (intrackability of the whole sequence, intrackability of adjacent frames and intrackability of single target in adjacent frames) are described in this paper. The simplicity of intrackability computation is also proposed. It is confirmed by experiments that four factors (target number in the scent, target resolution, target velocity and distinctiveness of tracking features) affect intrackability. And two approaches (automatic target combination and dynamic feature cascade in occlusion target tracking) of tracking occlusion targets are proposed according to intrackability theory.Due to the limitation of occlusion target tracking approach which is introduced ahead, another general framework of tracking occlusion target is proposed. In this framework, the theory of occlusion layers is introduced to represent target occlusion partial order which is determined using the appearance features and velocity features in target overlapping patches. Then the target non-occlusion parts can be obtained according to the target occlusion relation and target states. Occlusion target is described by the combination of appearance features and velocity features. The traditional mean shift tracking algorithm is improved to estimate target position using appearance histograms and velocity histograms. The probabilities of all the target scale changes at current time are predicted according to target scales changes at previous time.In this paper, Markov Chain Monte Carlo approach is used for estimating optimal states of occlusion targets, because three types of parameters (target position, target scale and target occlusion relations) which describe target states are correspondence with each other; the parameter space which includes discrete variables and continuous variables varies with the number of occlusion targets. During the process of estimating the optimal target states, in order to accelerate the algorithm converges, the state transaction functions are established for these parameters which include target positions; occlusion relations and target scales. In order to obtain the optimal states, these parameters are gradually adjusted one by one using sampling approach. When the tracking algorithm converges, the model state is optimal for occlusion target states.This tracking algorithm is test by several image sequences with different kinds of occlusion targets. From the tracking results, we can see that this tracking algorithm can well track occlusion targets.更多还原