【作者】 闻帆；

【导师】 王常虹；

【作者基本信息】 哈尔滨工业大学 ， 控制科学与工程， 2010， 博士

【摘要】 近年来,随着交通监控和交通管理的智能化水平不断提高,以视频图像处理、分析、理解为基础的视频监控技术越来越多地引起人们的重视。其中,智能交通系统中的交通检测与信息采集已经成为计算机视觉技术应用中的一项重要课题,而运动车辆的自动检测、分类、跟踪与冲突检测等则是其中最基础的部分。本论文针对上述问题进行了探索和研究,提出了新的方法,并通过实验证明了方法的有效性。本文的主要研究内容和学术上的主要贡献包含以下几个方面:首先,研究了运动目标检测问题。在分析交通场景阴影特性的基础上,提出了一种在马尔科夫随机场框架下的基于多特征信息融合的阴影检测算法。首先,利用基于高斯混合模型的背景减法提取前景运动目标,并利用改进的形态学滤波算法去除噪声干扰。在阴影检测过程中,综合利用前景像素和对应背景图像像素间的颜色、边界、纹理和时空一致性等特征信息来检测阴影,将各种特征信息集成到马尔可夫随机场能量函数中,并利用图割算法最小化马尔科夫随机场能量函数,在有效去除阴影的基础上得到最终的前景目标检测结果。其次,研究了运动目标分类问题。针对目标特征描述和分类器的选取两个影响目标分类精度的关键问题,提出了一种基于核主成分分析的梯度方向直方图特征算子,并构建了基于二叉决策树支持向量机的多类目标分类器。其中,在目标特征描述过程中,首先提取运动目标的梯度方向直方图,利用Mean-Shift聚类算法将特征向量聚类为若干个具有高度的内在相似性的子集。然后,将其映射到一个高维特征空间,进行线性主元分析,实现对目标特征的有效描述。在此基础上,利用构建的基于二叉决策树支持向量机分类器实现对场景中多类目标的准确分类。另外,为了实现对具有非线性、非高斯、多模态等运动特征的准确目标跟踪,提出了一种基于人工免疫算法的粒子滤波器。由于传统的粒子滤波算法在运算过程中会出现粒子退化现象,因此在重采样过程中引入人工免疫算法,通过对粒子的克隆选择使粒子样本集合保持一定的多样性,有效地避免了粒子的退化现象。针对复杂背景下视频目标跟踪的实时性和可靠性问题,设计了一种基于颜色特征和边缘梯度特征的自适应融合目标跟踪算法。在目标跟踪过程中,为了适应跟踪过程中目标与背景的变化,同时为了提高目标跟踪的鲁棒性,本文综合利用上述两种特征共同构建粒子滤波器的观测概率分布。准确的交通冲突检测结果是实现基于交通冲突技术的交叉口安全评价的前提。在完成轨迹聚类和冲突预测的基础上,提出了一种基于临界安全区域的交通冲突判别方法。首先,利用Mean-Shift算法对运动目标的初始轨迹集合进行聚类,得到若干个能够描述场景行为模式的轨迹子集合。然后,利用隐马尔科夫模型建立各个轨迹的模式类,并利用训练样本得到各个模型的相关参数。在冲突检测过程中,利用检测到的各个运动目标的部分轨迹序列信息进行冲突预测,然后综合利用目标间的距离、速度和行驶方向等检测数据判断冲突的发生与否和冲突的严重程度。更多还原

【Abstract】 With the improvement of traffic management and traffic monitoring intelligent level, video surveillance technologies based on video image processing, analysis and understanding have drawn increasing attention recently. Among them, traffic detection and informant collection in Intelligent Transportation System(ITS) area has become an important issue in the computer vision technology, and the most basic part is vehicles detection, classification, tracking and conflict detection automatically. In this paper, new methods are presented based on exploration and research on the issues mentioned above and experiments are conducted to demonstrate the effectiveness of the proposed methods.The major research contents and academic contributions of this dissertation include:First, the issue of moving object detection is studied. The algorithm for shadow detection based on multi-feature information fusion in the framework of Markov Random Field(MRF) is proposed based on the analysis of traffic scene shadow features. Background subtraction based on Gaussian mixture model is employed for foreground moving object extraction and the improved morphological filtering algorithm is used for eliminating noise. Shadow detection is accomplished based on the feature information, such as color, edge, texture, spatiotemporal coherence between the foreground pixels and the corresponding background image pixels. A variety of characteristic information is integrated into MRF energy function and graph cut algorithm is used for minimizing MRF energy function. Finally, foreground object detection results are achieved with effective shadow elimination.Next, the problem of moving object classification is solved. As the object characterization and classification are the key factors affecting classification accuracy, a histogram of oriented gradient operator based on kernel principal component analysis is proposed, and a binary decision tree based support vector machines for multi-class classification is constructed. Among them, in the object characterization process, the moving objects histogram of oriented gradient is obtained, and Mean-Shift clustering algorithm is used to cluster them into a number of feature vectors with a high degree of similarity among the internal subsets. Then, the subsets are mapped to a high dimensional feature space, and linear principal component analysis is employed to achieve an effective description of the object feature. With the discussion above, accurate object classification results could be obtained by the constructed support vector machine classifier based on binary decision tree.To achieve an accurate tracking of moving objects with non-linear, non-Gaussian, multi-modal motion characteristics, a particle filter based on artificial immune algorithm is proposed. As particle degradation phenomenon occurs during the calculation by traditional particle filter algorithm, artificial immune algorithm is introduced into the particle filter’s re-sampling process and colonel selection of the particles to maintain the particle sample collection diversity, which can alleviate the particle degradation phenomenon effectively. Aiming at the real-time and reliability problem of moving object tracking under complex background environment, we design an adaptive fusion object tracking algorithm based on the color feature and edge gradient feature. In the process of object tracking, the moving objects’ color feature and edge feature are fused into the observation probability distribution of the particle filter to compensate the changes of object and background, and improve tracking robustness.The accurate traffic conflict detection results are the precondition of intersection safety assessment based on traffic conflict technique. Upon completion of trajectory clustering and conflict prediction, a traffic conflict discrimination method based on critical security zone is proposed. First, Mean-Shift method is used to cluster moving objects’initial trajectory set, and a number of trajectory subsets which could represent scene behavior pattern are obtained. Then, each trajectory category is constructed by (Hidden Markov Models)HMM, and all models’parameters are obtained using training samples. In the conflict detection process, we use partial trajectories to predict conflict occurrence, and then the appearance and severity of a conflict are determined based on the features as following datum, distance, speed and running orientation.更多还原