摘要：

在视频数据处理过程中容易出现维数灾难的问题。为此,提出一种线性序列差异分析方法,对视频数据降维来进行人体行为识别。运用ViBe算法对视频帧进行背景减除操作获取行为区域,在该区域内提取稠密轨迹特征从而去除背景数据的干扰。使用Fisher Vector对特征编码后进行线性序列差异分析,采用动态线性规整算法计算序列类别间相似度,得到最小化类内残差和最大化类间残差的线性变换,将特征从高维空间投影至低维空间,降低特征维数。利用降维后的特征训练支持向量机,实现人体行为识别。在KTH数据集和UCF101数据集上进行数据仿真,结果表明,与主成分分析算法、线性判别分析法等相比,该方法可有效提高识别准确率。

关键词: 人体行为识别, 背景减除, 稠密轨迹, 线性序列差异分析, 降维

Abstract:

Aiming at the problem that dimensionality disaster easily occurs in the processing of dealing with video data,a dimension reduction method called Linear Sequence Discriminant Analysis(LSDA) is proposed for human action recognition.ViBe algorithm is used to subtract the backgrounds of video frames to get action areas,and dense trajectories are extracted in these areas to suppress the noise caused by camera movements.Fisher Vector is used to encode the features and linear sequence discriminant analysis is conducted on them,the sequence class separability is measured by dynamic time warping distance.In order to reduce the data dimension,a linear discriminative projection of the feature vectors in sequences is mapped to a lower-dimensional subspace by maximizing the between-class separability and minimizing the within-class separability.Support Vector Machine(SVM) is learned from the reduced dimension features,and then get the results of human action recognition.Simulation results on KTH datasets and UCF101 datasets show that compared with Principal Component Analysis(PCA),Linear Discriminant Analysis(LDA) and other dimension reduction methods,the proposed method can effectively improve the recognition accuracy.

Key words: human action recognition, background substraction, dense trajectories, Linear Sequence Discriminant Analysis(LSDA), dimension reduction

中图分类号: 

• TP391.41