监督型典型相关分析及其行为识别应用

SUPERVISED CANONICAL CORRELATION ANALYSIS WITH ITS APPLICATION ON ACTIVITY RECOGNITION

局部判别型典型相关分析CCA(LDCCA)是一种线性判别方法,只适用于处理线性可分的问题。为了更好地处理现实世界中存在的非线性现象,利用核技巧对LDCCA进行了核化,提出了一种新的有监督多模态识别方法即核化的局部判别型典型相关分析(KLDCCA)。LDCCA和KLDCCA引入了样本的类信息,充分考虑了同类样本之间的局部相关与不同类样本之间的局部相关关系及其对分类的影响,因此,提取的特征能够实现同类样本之间相关最大化,同时使得不同类样本之间相关最小化,这将有利于模式的分类。在人脸识别和简单行为识别上的应用表明,LDCCA和KLDCCA能有效地利用类信息和局部信息来提高分类性能。

Local discriminant canonical correlation analysis （LDCCA） is a linear discriminant method which is just used for linearly separable problems. In order to tackle the nonlinearly cases exist widely in the real world better, this paper proposes a new supervised multimodal recognition method called kernelised local discriminant canonical correlation analysis algorithm （abbreviated KLDCCA） based on the kernelisation of LDCCA with the help of kernel trick. LDCCA and KLDCCA introduce the class information of samples and fully consider the local correlation of both of the within-class sets and the between-class sets as well as their effect on classification, so features extracted by LDCCA can realise the maximisation of correlation of within-class sets and the minimisation of correlation of between-class sets both, which is good for classification of pattern. The application on facial recognition and simple activity recognition indicate that LDCCA and KLDCCA can effectively enhance the classification performance by using local and class information.