【作者】 韩旭；

【导师】 汤春明；

【作者基本信息】 哈尔滨工程大学 ， 通信与信息系统， 2014， 硕士

【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。更多还原

【Abstract】 With the continuous development of intelligent traffic surveillance technology, behavior analysis and recognition based on the trajectories of moving targets has become a hot topic,in which cluster analysis and anomaly detection are important research contents. Typical trajectory motion patterns of surveillance scene can be automatically obtained by clustering the trajectories of moving targets. Anomaly detection is aimed at detecting automatically abnormal behaviors in the surveillance scene and alarming timely, which is a key step to realize intelligent surveillance. This paper focuses on trajectory cluster analysis and anomaly detection in the field of intelligent traffic surveillance, and makes a thorough study of the problems in these two key technologies. Then practical solutions are proposed by taking full advantage of trajectory feature information. The main work and contributions of this paper are summarized as follows:In cluster analysis, traditional clustering algorithms only use single feature information for clustering, which reduces the accuracy of clustering. To solve this problem, hierarchical clustering algorithm based on multi-feature trajectory is proposed. This method uses Bhattacharyya distance and modified Hausdorff distance based segment interpolation (IMHD)to measure the similarity of motion direction and spatial location among trajectories respectively, and then extracts the trajectory motion patterns by the coarse-to-fine hierarchical clustering. To improve the efficiency of clustering, Laplacian mapping is introduced to reduce the computational complexity and to determine automatically the number of clusters in each layer of the agglomerative hierarchical clustering.In anomaly detection, a new description for abnormal behavior firstly is proposed.According to the different extent and nature that abnormal trajectories deviate from normal patterns, three common anomaly types are defined, which are starting point anomaly, global anomaly and local anomaly. The improved method effectively solves the problems that traditional description methods are not universal and define anomaly types vaguely. After that,anomaly detection method based on multi-feature trajectory is proposed to solve the problem that traditional methods only consider spatial location anomalies while ignoring direction anomalies, or can only detect abnormal trajectories with larger differences while ignoring local sub-segment abnormalities. This method firstly learns distribution patterns of starting location in surveillance scene by GMM model and establishes starting distribution model.Then the classifiers based on the position distance and direction distance are established by learning the location patterns and direction patterns of each normal trajectory classes after clustering,which take the moving window as a basic comparison unit. Finally at anomaly detection stage, combining with the anomaly types defined in this paper, the proposed abnormal detection algorithm based on multi-feature trajectory measures the differences between the tested trajectory and trajectory patterns in starting point distribution, position and direction, and judges whether it is abnormal trajectory. By sliding the moving window,the algorithm realizes the on-line detection for dynamically incremental trajectory data.Finally, clustering algorithm and anomaly detection method proposed in this paper are applied to vehicle trajectory data in real traffic scene. Experimental results show that the proposed methods can extract vehicle motion patterns of traffic scene quickly and accurately,and can automatically detect a variety of common abnormal behaviors. And the two methods are superior to traditional methods at clustering accuracy and abnormal recognition rate respectively.更多还原