面向复杂公共区域的群体聚集性计算方法研究

发布时间：2018-06-30 02:39

本文选题：群体运动 + 全局特征提取　；参考：《郑州大学》2017年硕士论文

【摘要】：随着城镇人口的增多和城市规模的不断扩大,发生在公共区域的群体聚集情形越来越常见,伴随而来发生在人群中的恐怖暴力和踩踏等群体性事件严重影响了社会的安全与稳定,针对此类场景中的群体聚集行为分析逐渐成为智能视频监控领域新的研究热点。由于公共区域内的目标对象类型较多、群体运动模式复杂多变,使得针对这些区域内的群体聚集状态分析变得非常困难。针对该问题,本文提出了一种基于运动目标轨迹信息的群体聚集状态计算方法。首先,采用了一种新的全局特征来表示公共区域内的群体运动,该特征可以充分描述目标兴趣点的时空和运动信息;其次,提出了一种对目标兴趣点先聚类分组再计算聚集性的策略,使得单个群体的聚集性计算更加一致有效;最后,提出了一种更为全面的群体聚集性描述子,用于对人群聚集状态进行详细描述,并以该描述子为基础,实现了群体运动状态演化分析与异常行为检测。在本文提出的群体聚集状态计算方法中,所采用的目标运动信息是利用KLT算法提取特征点的方法得到的,并基于这些特征点对人群密度做近似估计。这种方法能够处理摄像头距行人较远的场景,但当摄像头距行人较近时,这种方法就会存在较大误差。为对后续人群密度以及人群状态分析提供准确的数据,本文采用对行人进行跟踪的方法来提取这些信息。然而,由于监控视频中的目标对象姿态多变、稠密场景下个体间遮挡严重,传统的检测跟踪方法并不能很好的解决这些问题。因此,针对以上问题,本文提供了一种基于深度学习和卡尔曼滤波的行人目标跟踪算法。首先使用卷积神经网络对行人进行检测;然后针对检测过程中出现的漏检现象,采用卡尔曼滤波算法来预测行人当前的位置,最后使用匈牙利分配算法关联视频相邻帧中的行人,从而实时的抽取行人运动信息。实验结果表明,本文所提出的群体运动特征提取方法能够很好地应用于中高密度场景下的人群运动特征计算;结合这些精确的人群运动特征,本文提出的群体聚集性描述子能够应用于各种复杂公共区域内的群体行为分析,可为公共安全管理提供真实可靠的参考。
[Abstract]:With the increase of urban population and the continuous expansion of urban scale, it is becoming more and more common for groups to gather in public areas. The terrorist violence and stampede in the crowd have seriously affected the security and stability of the society. The analysis of the collective behavior in this kind of scene has gradually become a new research hotspot in the field of intelligent video surveillance. Because of the variety of target object types and the complex and changeable movement patterns in common areas, it is very difficult to analyze the aggregation state of groups in these areas. In order to solve this problem, this paper presents a method to calculate the cluster state based on the trajectory information of moving targets. First of all, a new global feature is used to represent the group movement in the common area, which can fully describe the space-time and motion information of the point of interest of the target. In this paper, a strategy of clustering and then computing aggregation is proposed, which makes the aggregation calculation of a single population more consistent and effective. Finally, a more comprehensive cluster descriptor is proposed. It is used to describe the state of crowd aggregation in detail, and based on the descriptor, the evolution analysis of group motion state and the detection of abnormal behavior are realized. In the proposed method, the target motion information is obtained by using KLT algorithm to extract feature points, and the population density is estimated approximately based on these feature points. This method can deal with the scene where the camera is far from the pedestrian, but when the camera is close to the pedestrian, there will be a big error. In order to provide accurate data for the subsequent population density and crowd state analysis, this paper uses the method of pedestrian tracking to extract the information. However the traditional detection and tracking methods can not solve these problems because of the changeable attitude of the target object in the surveillance video and the serious occlusion between individuals in dense scenes. Therefore, a pedestrian target tracking algorithm based on depth learning and Kalman filter is proposed. Firstly, we use convolutional neural network to detect pedestrians, and then we use Kalman filter algorithm to predict the current position of pedestrians. Finally, the Hungarian assignment algorithm is used to correlate the pedestrians in the adjacent frames of the video, so that the pedestrian motion information can be extracted in real time. The experimental results show that the method proposed in this paper can be well applied to the calculation of crowd motion characteristics in medium and high density scenes, and combined with these accurate crowd motion features, The group aggregation descriptor proposed in this paper can be applied to the analysis of group behavior in various complex public areas, and can provide a true and reliable reference for the management of public safety.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】