摘要
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,与线性近似稀疏表示(LASRC)快速分类算法结合用于大数据的行为识别。利用LASRC的残差计算规律设计OP-LASRC实现监督降维:在追求一个线性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,保留样本的类别特征。对降维后的行为数据用LASRC算法分类:用L2范数估算稀疏系数,取前k个最大的稀疏系数对应的训练样本,用L1范数和残差计算得到识别结果。在KTH行为数据库上的实验表明,OP-LASRC降维后,LASRC在分类时识别率高达96.5%,执行时间比同类算法短,抗噪声能力强,证明了OP-LASRC的高效和强鲁棒性,能完美匹配LASRC用于大数据的行为识别。
In view of the low real-time performance and low recognition rate in human behavior recognition of large data, this paper proposed a supervised dimensionality reduction algorithm. It combined with linearly approximated spare representation based classification (LASCR) algorithm to use in large data of human behavior recognition. Firstly, this algorithn designed LASCR residual calculation rules OP-LASRC algorithm, which realized supervised dimensionality reduction: the orthogonal projection reduced the training samples of the between-class reconstruction residual and increased with-class reconstruction re- sidual, so as to preserve class features of training samples. Then, it used LASCR to classify behavior data: estimated the sparse coefficients with the L2 norm, selected training samples corresponding to the k of the largest sparse coefficients, after used L1 norm and residuals to obtain recognition results. Experimental results show that the OP-LASRC matched LASRC in the classification not only reaches recognition rate of 96.5 % on the KTH action database, execution time is shorter than the other algorithms of the same kind, but also ensures the robustness. It proves the OP-LASRC can perfect match LASCR algorithm to use for recognition behavior of big data.
出处
《计算机应用研究》
CSCD
北大核心
2017年第11期3477-3481,3485,共6页
Application Research of Computers
基金
国家自然科学基金资助项目(41075019)
关键词
稀疏表示
监督降维
优化投影
线性近似
行为识别
sparse representation
supervise dimension reduction
optimize projection
linear approximation
action recognition