摘要
针对场景密集人群移动过程中存在遮挡和倒影问题,提出一种新的谱聚类算法用于人群聚类。首先构造一种新的邻接矩阵作为谱聚类的输入参数,然后再构造一种新的拉普拉斯矩阵,通过选取拉普拉斯矩阵的四个最小特征值组成特征向量。采用K-means算法对特征向量进行聚类得到人群聚类指数,并将聚类指数映射至图像。通过在CCD、CMD、MPT国际公开数据集上进行实验,上述算法较其它其他人群聚类算法得到更高的聚类纯度(Purity)、FM(F-Measure)值、AUC值和更高的检测率,说明了本文提出的基于改进谱聚类算法的有效性。
A novel spectral clustering algorithm is proposed to solve the problem of occlusion and reflection in the movement of dense crowd. Firstly, a new adjacency matrix was constructed as the input parameter of spectral clustering. Then, a new Laplacian matrix wes constructed, and the eigenvector was composed of four minimum eigenvalues of the Laplace matrix. Finally, k-means algorithm was used to cluster the feature vectors to get the clustering index of people group, and the clustering index was mapped to the image. Through experiments on CCD, CMD and MPT international open data sets, compared with other group clustering algorithms, this algorithm obtains higher clustering purity, FM value, AUC value and higher detection rate;the experimental results prove the effectiveness of the improved spectral clustering algorithm proposed in this paper.
作者
杨玉成
张乾
邵定琴
岳诗琴
YANG Yu-cheng;ZHANG Qian;SHAO Ding-qin;YUE Shi-qin(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang Guizhou 550025,China;Pattern Recognition and Intelligent Systems Key Laboratory of Guizhou,Guiyang Guizhou 550025,China)
出处
《计算机仿真》
北大核心
2021年第6期450-454,共5页
Computer Simulation
基金
国家自然科学基金项目(61802082,61762020)
贵州省教育厅自然科学基金项目(黔教合KY字[2017]129,黔教合KY字[2018]018)。
关键词
行为检测
人群移动
谱聚类
拉普拉斯矩阵
Behavior detection
Crowd movement
Spectral clustering
Laplacian matrix
