一种基于局部拓扑与l_(1/2)范数的解析字典分类的人群事件检测  

Abnormal event detection based on local topology and l_(1/2) norm regularize

作　　者：禹青 陈恳[1] 李萌[1] 李斐[1] YU Qing;CHEN Ken;LI Meng;LI Fei(Institute of Communication Technology, Ningbo University)

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315211

出　　处：《电信科学》2018年第10期65-71,共7页Telecommunications Science

基　　金：国家自然科学基金资助项目(No.60972063);宁波市自然科学基金资助项目(No.2014A610065);宁波大学科研基金(理)/学科资助项目(No.XKXL1308)~~

摘　　要：所提方案在传统解析字典算法基础上,加入局部拓扑项用以描述数据之间的结构信息,同时用l_(1/2)范数代替l1范数作为稀疏约束,从而提高表示系数的稀疏度。在特征提取上,融合了包含丰富运动信息的相互作用力直方图与包含纹理信息的梯度方向直方图,然后用改进的字典对特征数据进行训练,最后通过计算测试样本在该字典下的重构误差来判断测试样本是否为异常样本。在标准行为库UMN(University of Minnesota)数据库上进行的实验证实了算法具有较高的性能。与传统的算法相比,提出的改进的解析字典分类算法在针对人群异常事件中取得了更为有效的检测。A new dictionary learning method was proposed by introducing a local topology term to describe structural information of video events and using the l1/2 norm as the sparsity constraint to the representation coefficients based on the traditional analysis dictionary learning method. In feature extraction, a histogram of interaction force（HOIF） containing rich motion information and a histogram of oriented gradient（HOG） containing texture information were merged. Then, the improved dictionary was used to train the feature data. Finally, the reconstruction error of the testing sample under the dictionary was used to determine whether the testing sample was an abnormal sample. Experiments on UMN show the high performance of the algorithm. Compared with the state-of-the-art algorithms, the analysis dictionary classification algorithm based on local topology and l1/2 norm has made more effective detection on the abnormal events in the crowd.

关 键 词：解析字典 局部拓扑项 相互作用力直方图 梯度方向直方图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]