基于车辆轨迹多特征的聚类分析及异常检测方法的研究

发布时间：2018-05-21 05:11

本文选题：智能交通监控 + 轨迹多特征　；参考：《哈尔滨工程大学》2014年硕士论文

【摘要】：随着智能交通监控技术的不断发展,基于运动目标轨迹的行为分析和识别已成为研究热点,其中聚类分析和异常检测是研究的重点内容。通过对运动目标的轨迹进行聚类,可以自动的获取监控场景的典型轨迹运动模式并了解场景结构;而异常检测的目标是能实时的自动的检测出监控场景中的异常行为,并及时的报警异常,是实现智能化监控的关键步骤。本文对智能交通监控领域中的轨迹聚类分析和异常检测这两个关键技术中存在的问题进行了深入的研究,充分利用了轨迹的不同特征信息,提出了切实可行的改进方法,主要工作体现在以下几个方面:在聚类分析方面,针对传统聚类方法只利用轨迹的单一特征信息进行聚类,聚类准确率低的问题,提出了基于轨迹多特征的分层聚类算法。该方法分别采用Bhattacharyya距离和基于线段插值的改进Hausdorff距离衡量轨迹间运动方向和空间位置的相似度,通过由粗到细的分层聚类来提取轨迹运动模式。为了提高聚类的效率,在每层的凝聚层次聚类中引入Laplacian映射以降低计算复杂度并同时自动确定每层的聚类数目。在异常检测方面,首先对异常行为进行了全新的描述。根据异常轨迹偏离正常模式的程度和性质的不同,定义了三种常见的异常类型,分别为起点异常、全局异常和局部异常,有效解决了传统的异常描述方法通用性不强、异常类型定义模糊等问题。然后针对传统的异常检测方法只考虑了轨迹的空间位置异常而忽略方向异常,或只能粗略检测差异较大的轨迹异常而忽略轨迹局部子段异常等问题,提出了基于轨迹多特征的在线异常检测方法。该方法先通过GMM模型学习监控场景的轨迹起点位置分布模式,建立轨迹起点分布模型;再以移动窗作为基本比较单元,学习聚类后的每个正常轨迹参考类的空间位置模式和运动方向模式,建立基于位置距离和方向距离的分类器。最后在异常检测阶段,结合本文定义的异常类型,通过提出的在线多特征异常检测算法从起点分布、空间位置和运动方向三个层次来衡量待测轨迹和正常轨迹模式之间的差异,判断轨迹是否异常;并通过滑动移动窗口的方式,实现了对动态递增轨迹数据的在线检测。最后,将本文提出的聚类算法和异常检测方法应用于真实交通场景的车辆轨迹数据中。实验结果表明,本文的方法能快速准确的提取交通场景的车辆运动模式并能自动检测出各种常见的交通异常行为,而且两种方法分别在聚类准确率和异常识别率上更优于传统方法。
[Abstract]:With the continuous development of intelligent traffic monitoring technology, behavior analysis and recognition based on moving target trajectory has become a hot research topic, among which clustering analysis and anomaly detection are the focus of the research. By clustering the trajectory of moving objects, we can automatically obtain the typical trajectory motion pattern and understand the scene structure; and the object of anomaly detection is to detect the abnormal behavior in the monitoring scene in real time and automatically. And timely alarm abnormal, is a key step to achieve intelligent monitoring. In this paper, the problems of trajectory clustering analysis and anomaly detection in intelligent traffic monitoring field are deeply studied, the different characteristic information of trajectory is fully utilized, and a feasible improvement method is put forward. The main work is as follows: in the aspect of clustering analysis, aiming at the problem that the traditional clustering method only uses the single feature information of the trajectory to cluster, and the accuracy of clustering is low, a hierarchical clustering algorithm based on multiple locus features is proposed. In this method, Bhattacharyya distance and improved Hausdorff distance based on line segment interpolation are used to measure the similarity between trajectory direction and space position, and the trajectory motion pattern is extracted by hierarchical clustering from coarse to fine. In order to improve the efficiency of clustering, Laplacian mapping is introduced into the cluster of condensed layers in each layer to reduce the computational complexity and determine the number of clusters in each layer automatically at the same time. In the aspect of anomaly detection, the abnormal behavior is described completely new. According to the degree and nature of abnormal trajectory deviating from normal mode, three common types of anomaly are defined, which are starting point anomaly, global anomaly and local anomaly. The definition of exception type is vague and so on. Then the traditional anomaly detection method only considers the space position anomaly of the trajectory and neglects the direction anomaly, or only roughly detects the track anomaly with great difference and neglects the local subsegment anomaly of the trajectory, and so on. An online anomaly detection method based on multiple locus features is proposed. In this method, first of all, the GMM model is used to study the distribution pattern of the locus starting point of the scene, and then the moving window is used as the basic comparison unit. A classifier based on position distance and direction distance is established by studying the spatial position pattern and moving direction pattern of each normal trajectory reference class after clustering. Finally, in the phase of anomaly detection, combined with the anomaly types defined in this paper, the online multi-feature anomaly detection algorithm is proposed to measure the difference between the trajectory to be tested and the normal trajectory pattern from three levels: starting point distribution, space position and movement direction. To judge whether the trajectory is abnormal or not, and to realize the on-line detection of dynamic incremental trajectory data by sliding moving window. Finally, the proposed clustering algorithm and anomaly detection method are applied to the vehicle trajectory data of real traffic scene. The experimental results show that the proposed method can quickly and accurately extract the vehicle motion patterns of traffic scenes and detect all kinds of common abnormal traffic behaviors automatically. Moreover, the two methods are better than the traditional methods in clustering accuracy and anomaly recognition rate respectively.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495;TP311.13

【参考文献】