高密度场景下人群中组群的识别与分析

发布时间：2018-03-16 14:34

本文选题：组群识别　切入点：人群行为分析　出处：《北京交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：基于组群层次的人群行为分析是近年来一种新而有效的研究方法,在视频监控和人群理解等领域有很多重要的应用。然而,组群作为人群的基本组成单位,对人群有很大的影响,对组群的深入研究具有很大的意义。本文的工作重点在于给出一种高密度场景下人群中组群的识别方法,对组群的一些通用特征进行研究并给出定量描述,最后分析组群中的重要人物,构建组群的三层结构模型。论文前半部分提出了一种无需人工参与即能快速自动进行组群识别的方法。首先,使用自动而快速的密度聚类算法(Automaticand fast density clustering,AFDC)寻找组群的核心,然后通过组群成员的相关邻居不变性(Coherent neighbor invariance,CNI)进行融合。相较于传统方法,该方法既考虑了组群在单帧下的空间和运动方向特征,又考虑了多帧下时空上的不变特征,因此具有高度的鲁棒性和广泛的适用性。基于数百个公共场景下视频片断的实验效果表明,对于形状和密度不断变化的组群,这种组群识别方法的效果也很好。由于避免了大量的迭代计算,在保证我们的组群识别方法取得优良的识别效果的基础上,算法的性能也得到了很大的改善。对组群的统计实验结果显示组群数量呈现一种幂律分布(Power-law distribution),这对人群仿真建模,场景理解等领域有着重要的意义。组群中个体的相对重要性研究也引发了广泛的关注。论文接下来的工作中采用了 45维的向量表示个体特征,并采用支持向量回归算法(Support vector regression,SVR),建立了组群成员重要性预测模型,判定组群成员中的重要人员。由重要成员,核心成员,相关人员三层结构的人员就构成了组群的三层结构模型。论文最后将组群视作无向图,用顶点表示人群中的个体,用边表示两两个体之间的距离权重。对组群轨迹使用马尔科夫链进行建模,求取组群先验(Prior),然后以此为基础综合分析组群的地理位置结构,组群的内部拓扑结构,聚集程度以及组群间的相互作用等信息,并给出聚集度,稳定度,均匀度,冲突度等几种基础而通用的组群特征的设计和定量的定义。这些组群特征的场景无关性很好的说明了其广泛的适用性,并且能够有效应用到人群仿真建模,视频分类,人群状态监控,人群视频检索等领域。对高密度场景的理解,人群行为分析等也将产生重要的影响。
[Abstract]:Crowd behavior analysis based on cluster level is a new and effective research method in recent years, which has many important applications in the field of video surveillance and crowd understanding. It has great influence on the crowd, and has great significance for the further study of the cluster. The main work of this paper is to give a method of group identification in the high density scene. Some common features of the cluster are studied and quantitatively described. Finally, the important persons in the cluster are analyzed. In the first half of this paper, a fast and automatic cluster identification method is proposed in the first half of the paper. Firstly, the automatic and fast density clustering algorithm is used to find the core of the cluster. Then the fusion is carried out by using the relative neighbor invariance of cluster members, which takes into account not only the spatial and motion direction characteristics of the cluster in a single frame, but also the invariant features in space and time under multiple frames, compared with the traditional method. Therefore, it has a high degree of robustness and wide applicability. Based on the experimental results of hundreds of common scenes, the experimental results show that, for clusters with changing shapes and densities, The effect of this method is also very good. Because of avoiding a large number of iterative calculations, we can ensure that our cluster recognition method has good recognition effect. The performance of the algorithm has also been greatly improved. The results of the statistical experiments on the clusters show that the number of clusters presents a power law distribution, which is used to model the crowd simulation. The research on the relative importance of individuals in the cluster has attracted extensive attention. In the following work, 45 dimensional vectors are used to represent individual characteristics. The support vector regression algorithm is used to predict the importance of cluster members. In the end, the group is regarded as an undirected graph, and the vertex is used to represent the individual in the crowd. The distance weight between two individuals is represented by edges. Markov chain is used to model the group locus, and a priori priorn is obtained. Based on this, the geographical position structure of the group and the internal topological structure of the group are analyzed synthetically. The degree of aggregation and the interaction between clusters are given, and the degree of aggregation, stability and uniformity are given. The design and quantitative definition of several basic and general cluster features, such as conflict degree, are used to illustrate their wide applicability, and can be effectively applied to crowd simulation modeling, video classification, and so on. Crowd state monitoring, crowd video retrieval and other fields. The understanding of high-density scenes, crowd behavior analysis will also have an important impact.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】