空间序列低秩稀疏子空间聚类算法

Low Rank and Sparse Subspace Clustering Algorithm for Spatial Sequence

研究序列数据的子空间聚类问题,具体来说,给定从一组序列子空间中提取的数据,任务是将这些数据划分为不同的不相交组。基于表示的子空间聚类算法,如SSC和LRR算法,很好地解决了高维数据的聚类问题,但是,这类算法是针对一般数据集进行开发的,并没有考虑序列数据的特性,即相邻帧序列的样本具有一定的相似性。针对这一问题,提出了一种新的低秩稀疏空间子空间聚类方法(Low Rank and Sparse Spatial Subspace Clustering for Sequential Data,LRS3C)。该算法寻找序列数据矩阵的稀疏低秩表示,并根据序列数据的特性,在目标函数中引入一个惩罚项来加强近邻数据样本的相似性。提出的LRS3C算法充分利用空间序列数据的时空信息,提高了聚类的准确率。在人工数据集、视频序列数据集和人脸图像数据集上的实验表明:提出的方法LRS3C与传统子空间聚类算法相比具有较好的性能。

In this paper,the problem of subspace clustering for sequential data is studied.Specifically,given the data extracted from a set of sequential subspaces,the task is to divide data into different disjoint groups.Representation-based subspace clustering algorithms,such as SSC and LRR,can effectively solve the clustering problem of high-dimensional data.However,these algorithms are developed for general data sets without considering the characteristics of sequential data,that is,the samples of adjacent frame sequences have certain similarity.To deal with the problem,a novel method of low rank and sparse spatial subspace clustering(LRS3C)is proposed.The algorithm finds the sparse low rank representation of sequential data matrix,a penalty term is introduced into the objective function to enhance the similarity of neighbor data samples according to the characteristics of the sequential data.The proposed LRS3C algorithm makes full use of the spatiotemporal information of spatial sequential data and improves the accuracy of clustering.Experiments on artificial data sets,video sequence data sets and face image data sets show that the proposed LRS3C method has better performance than the traditional subspace clustering algorithm.