一种监督降维的OP-LASRC算法在行为识别中的应用*

针对大数据的人体行为识别时实时性差和识别率低的问题,提出了优化投影对线性近似稀疏表示分类(OP-LASRC)的监督降维算法。OP-LASRC将高维的行为数据优化投影到低维空间,与线性近似稀疏表示(LASCR)快速分类算法相结合应用大数据的人体行为识别。首先利用LASCR的残差计算规律设计OP-LASRC算法,实现监督降维；利用线性正交投影缩减高维数据的维度,投影时减小训练样本的本类重构残差及增大类间重构残差,从而保留训练样本的类别特征。然后,对降维后的行为数据,利用LASCR算法进行分类；用L2范数估算稀疏系数,选出前k个最大的稀疏系数对应的训练样本,缩减训练样本库后用L1范数最小化和残差最小化计算得到识别结果,从识别率、鲁棒性、执行时间三个方评价此方法,在KTH行为数据库上进行实验测试。实验表明：OP-LASRC监督降维后,LASRC在分类时不仅识别率高达96.5%,执行时间比同类算法短,而且保证了强鲁棒性,证明了OP-LASRC能完美匹配LASCR算法用于行为识别,这为大数据的行为识别提供了一种新的思路。

A supervised dimension reduction algorithm called OP-LASRC andits applications to action recognition

In view of the low real-time performance and low recognition rate in human behavior recognition of large data, a supervised dimensionality reduction algorithm is proposed: optimized projection for the linear approximate sparse representation classification (OP-LASRC). OP-LASRC optimized the high dimensional behavior of the data projection to the low dimensional space, and combines with linearly approximated spare representation based classification (LASCR) algorithm is used in large data of human behavior recognition. Firstly, LASCR residual calculation rules is designed OP-LASRC algorithm, which realizes supervised dimensionality reduction; the orthogonal projection of linear reduced the high-dimensional data, projection reduces the training samples of the between-class reconstruction residual and increase with-class reconstruction residual, so as to preserve class features of training samples. Then, classifying the dimensionality behavior data using LASCR algorithm ; the L2 norm estimates the sparse coefficients , selecting k a largest sparse coefficients corresponding to the training samples, after the smaller training sample library is used for L1 norm minimization and minimization residuals obtaining recognition results, the recognition rate, robustness, executive time three aspects of this method as evaluation indexes and on the KTH action database for testing. Experimental results show that the OP-LASRC matched LASRC in the classification not only reached recognition rate of 96.5%, execution time shorter than the other algorithms of the same kind , but also ensured the robustness. That proves the OP-LASRC can perfect matching LASCR algorithm is used for recognition behavior, providing a new way of thinking to the behavior recognition of big data.