人群密集场景中运动模式识别  

Motion Patterns Learning in Crowded Scenes

作　　者：康永欣 袁丽英[1] KANG Yong-xin;YUAN Li-ying(School of Automation,Harbin University of Science and Technology,Harbin 150080,China)

机构地区：[1]哈尔滨理工大学自动化学院

出　　处：《哈尔滨理工大学学报》2019年第3期74-81,共8页Journal of Harbin University of Science and Technology

基　　金：国家自然科学基金(61305001)

摘　　要：针对从人群密集场景中识别运动模式的问题,提出了距离依赖中餐馆连锁店过程混合模型。该模型是一种引入依赖关系的层次化非参贝叶斯模型,能够通过引进独立于数据观测值的距离依赖信息,准确建模自然分组的数据,从中挖掘共享的数据模式。给出了模型的建立过程,并通过吉布斯采样的方法对模型进行求解,同时展示了相关的实验结果。通过对纽约广场火车站监控场景数据集中47 866条片段轨迹的建模分析,证明了模型可以自动确定场景中运动模式的个数,从不完整的轨迹中以98%的正确度学习并表达运动模式,并且能够在不同的运动模式之间共享公共的子模式。To address the motion patterns learning task in crowded scenes,we propose a novel Distance Dependent Chinese Restaurant Franchise(DDCRF)mixture model,which is a hierarchical nonparametric Bayesian model based on dependencies.DDCRF can learn the latent patterns accurately by introducing the distance information which is dependent of the observations of data points.We detail the generative process and Gibbs sampling process of DDCRF,and then the results of experiments are shown.An extensive evaluation is performed on the dataset including 47,866 tracklets collected from the crowded New York Grand Central station,indicating that our algorithm has the following advantages:deducing the number of latent motion patterns automatically,learning motion patterns precisely from these tracklets,and sharing constituent parts and sub-patterns among different motion patterns.

关 键 词：运动模式识别 人群密集场景 非参贝叶斯模型 吉布斯采样 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]