CASIA OpenIR  > 毕业生  > 硕士学位论文
监控场景下群体行为分析研究; 监控场景下群体行为分析研究
赵炜琪
学位类型工程硕士 ; 工程硕士
导师黄凯奇 ; 黄凯奇 ; 张彰 ; 张彰
2017-05-25 ; 2017-05-25
学位授予单位中国科学院研究生院 ; 中国科学院研究生院
学位授予地点北京 ; 北京
关键词时空相似度 时空相似度 组群检测 组群检测 语义区域 语义区域 层级聚类 层级聚类
其他摘要
受实际应用如群体计数、群体密度估计和群体行为分析等有关公共安全需求的推动,近年来较多研究者投入到群体行为分析研究领域,针对不同问题提出了相应的解决方案。启发于格式塔聚类准则对群体形成过程的研究和描述,我们针对群体形成过程提出三个聚类优先描述子,并基于这三个聚类优先描述子将运动轨迹之间的时序信息与空间关系进行融合,定义了运动轨迹之间统一的时空相似度度量模型,从而在更长时序范围对群体动态行为进行挖掘和描述。为了较好地对我们的方法在更长时序段内的组群检测结果进行评测,我们通过园区内监控录像建立了拥有较长时序并逐帧标定真实运动轨迹标签的群体行为分析数据库。本文的工作主要有:
∙ 基于各个群体分析数据库主要解决问题的不同,对现有群体分析数据库进行分类介绍和分析,并在相应群体分析数据库上通过试验表明,目前的群体分析数据库对于验证基于较长时序运动轨迹时空关系挖掘动态组群的方法并不适用,因此我们建立了基于较长时序的CASIA Crowd群体行为分析数据库,并通过在我们新建立的群体分析数据库上进行的实验证明新建立的数据库更适用于较长时序的群体行为分析;
∙ 基于格式塔理论关于群体形成的准则和规律,我们设计了三个聚类优先描述子对运动轨迹的时空关系进行描述。这些描述子将存在于不同时序段内运动轨迹之间的时序关系与其空间近邻关系相融合,从而在更长的时序上建立运动轨迹之间统一的时空相似度模型,用于对运动轨迹之间的动态信息进行挖掘和表达,从而对群体的动态行为进行更全面的建模和描述;另外为了使运动轨迹的聚类更加合理和接近真实的群体动态行为,我们引入了时序窗口,并使用该时序窗口沿时间轴将轨迹的时序空间划分为不同的区间进行运动轨迹分析;
∙ 基于我们定义的运动轨迹之间统一的时空相似度模型,我们选用层级聚类方法对运动轨迹进行表达。不同于目前的群体分析方法需要人为提前设定聚类数目参数,我们的方法对于不同群体动态场景可以根据群体自身紧密程度自动确定聚类数目,这使得我们的方法有更好的扩展性。为了提高算法效率,我们的方法在聚类过程中提取代表性运动轨迹代表该类别的所有运动轨迹,并将该代表性轨迹用于后续聚类过程和运动路径分析。

;
受实际应用如群体计数、群体密度估计和群体行为分析等有关公共安全需求的推动,近年来较多研究者投入到群体行为分析研究领域,针对不同问题提出了相应的解决方案。启发于格式塔聚类准则对群体形成过程的研究和描述,我们针对群体形成过程提出三个聚类优先描述子,并基于这三个聚类优先描述子将运动轨迹之间的时序信息与空间关系进行融合,定义了运动轨迹之间统一的时空相似度度量模型,从而在更长时序范围对群体动态行为进行挖掘和描述。为了较好地对我们的方法在更长时序段内的组群检测结果进行评测,我们通过园区内监控录像建立了拥有较长时序并逐帧标定真实运动轨迹标签的群体行为分析数据库。本文的工作主要有:
∙ 基于各个群体分析数据库主要解决问题的不同,对现有群体分析数据库进行分类介绍和分析,并在相应群体分析数据库上通过试验表明,目前的群体分析数据库对于验证基于较长时序运动轨迹时空关系挖掘动态组群的方法并不适用,因此我们建立了基于较长时序的CASIA Crowd群体行为分析数据库,并通过在我们新建立的群体分析数据库上进行的实验证明新建立的数据库更适用于较长时序的群体行为分析;
∙ 基于格式塔理论关于群体形成的准则和规律,我们设计了三个聚类优先描述子对运动轨迹的时空关系进行描述。这些描述子将存在于不同时序段内运动轨迹之间的时序关系与其空间近邻关系相融合,从而在更长的时序上建立运动轨迹之间统一的时空相似度模型,用于对运动轨迹之间的动态信息进行挖掘和表达,从而对群体的动态行为进行更全面的建模和描述;另外为了使运动轨迹的聚类更加合理和接近真实的群体动态行为,我们引入了时序窗口,并使用该时序窗口沿时间轴将轨迹的时序空间划分为不同的区间进行运动轨迹分析;
∙ 基于我们定义的运动轨迹之间统一的时空相似度模型,我们选用层级聚类方法对运动轨迹进行表达。不同于目前的群体分析方法需要人为提前设定聚类数目参数,我们的方法对于不同群体动态场景可以根据群体自身紧密程度自动确定聚类数目,这使得我们的方法有更好的扩展性。为了提高算法效率,我们的方法在聚类过程中提取代表性运动轨迹代表该类别的所有运动轨迹,并将该代表性轨迹用于后续聚类过程和运动路径分析。

;
Crowded scene analysis is a popular research topic due to its great application potentials,
such as intelligent video surveillance, crowd density estimation and public security,and relevant methods are proposed according to various tasks. Inspired by theGestalt laws of grouping, which regarded as a set of principles accounting for the observation that humans naturally perceive complex scenes as organized patterns and objects, we propose three priors to define a unified spatial-temporal similarity metric to fuse the spatial correlation and temporal information of pairs of tracklets from di erent crowd groups, which preferable conveys the long-term dynamic behaviors of crowd. In order to verify our method in longer time duration video sequences in the crowded scene, we construct a new crowd database with various crowd densities in real garden scenes, and label the ground truth frame by frame. The main contribution of this paper as follows:
∙ Based on the di erent tasks that crowd databases focus on, this thesis gives detailed
introduction and analysis in di erent classes on existing crowd databases. Extensive experiments demonstrate that the existing crowd databases are not suitable to evaluate the long-term based method proposed in this paper, so it is necessary to build a new long-term based crowd database, and the experimental results on our new database verify that our new dataset is more suitable for long-term based crowd analysis methods;
∙ According to the Gestalt laws of grouping, we propose three priors to define a unified spatial-temporal similarity metric to measure the similarities of pairs of tracklets, which fuse the spatial correlation and temporal information of trackletsin longer temporal duration to better express the dynamic behaviors of crowd. In order to make the clustering process more reasonable and closer to the real crowd behavior, we apply the temporal window to divide the time axis into several temporal intervals, and cluster tracklets in each temporal window respectively;
∙ We apply the hierarchical clustering method to cluster tracklets based on the unified spatial-temporal similarity measurement. Di erent from other clustering approaches need to predefine the cluster number manually, our method can automatically decide when the clustering process stop and discover the optimal number of groups based on their intra-group tightness and inter-group di erence, which makes our method more extensible to di erent crowd scenes. In addition, in order to improve the eciency, we extract the representative tracklets from each cluster, and these representative tracklets will be used to higher level cluster procedure and path modeling.
;
Crowded scene analysis is a popular research topic due to its great application potentials,
such as intelligent video surveillance, crowd density estimation and public security,and relevant methods are proposed according to various tasks. Inspired by theGestalt laws of grouping, which regarded as a set of principles accounting for the observation that humans naturally perceive complex scenes as organized patterns and objects, we propose three priors to define a unified spatial-temporal similarity metric to fuse the spatial correlation and temporal information of pairs of tracklets from di erent crowd groups, which preferable conveys the long-term dynamic behaviors of crowd. In order to verify our method in longer time duration video sequences in the crowded scene, we construct a new crowd database with various crowd densities in real garden scenes, and label the ground truth frame by frame. The main contribution of this paper as follows:
∙ Based on the di erent tasks that crowd databases focus on, this thesis gives detailed
introduction and analysis in di erent classes on existing crowd databases. Extensive experiments demonstrate that the existing crowd databases are not suitable to evaluate the long-term based method proposed in this paper, so it is necessary to build a new long-term based crowd database, and the experimental results on our new database verify that our new dataset is more suitable for long-term based crowd analysis methods;
∙ According to the Gestalt laws of grouping, we propose three priors to define a unified spatial-temporal similarity metric to measure the similarities of pairs of tracklets, which fuse the spatial correlation and temporal information of trackletsin longer temporal duration to better express the dynamic behaviors of crowd. In order to make the clustering process more reasonable and closer to the real crowd behavior, we apply the temporal window to divide the time axis into several temporal intervals, and cluster tracklets in each temporal window respectively;
∙ We apply the hierarchical clustering method to cluster tracklets based on the unified spatial-temporal similarity measurement. Di erent from other clustering approaches need to predefine the cluster number manually, our method can automatically decide when the clustering process stop and discover the optimal number of groups based on their intra-group tightness and inter-group di erence, which makes our method more extensible to di erent crowd scenes. In addition, in order to improve the eciency, we extract the representative tracklets from each cluster, and these representative tracklets will be used to higher level cluster procedure and path modeling.
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14667
专题毕业生_硕士学位论文
作者单位中国科学院自动化研究所智能感知与计算研究中心
推荐引用方式
GB/T 7714
赵炜琪. 监控场景下群体行为分析研究, 监控场景下群体行为分析研究[D]. 北京, 北京. 中国科学院研究生院, 中国科学院研究生院,2017, 2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
赵炜琪-master_thesis.pd(8808KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[赵炜琪]的文章
百度学术
百度学术中相似的文章
[赵炜琪]的文章
必应学术
必应学术中相似的文章
[赵炜琪]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。