数据内在结构驱动的大间隔特征提取算法及其应用研究

项目名称： 数据内在结构驱动的大间隔特征提取算法及其应用研究

项目编号： No.61203244

项目类型： 青年科学基金项目

立项/批准年度： 2013

项目学科： 自动化学科

项目作者： 陈小波

作者单位： 江苏大学

项目金额： 24万元

中文摘要： 从数据中挖掘最有鉴别能力的特征是模式识别的一个基本问题。本项目旨在针对高维小样本数据和流形分布数据，借鉴最优化领域前沿方法，结合稀疏学习理论，研究大间隔特征提取模型和算法，并应用于驾驶员头部姿态估计。面向高维小样本数据，从特征空间分解、大间隔模型构建和参数优化等角度，研究能利用所有鉴别信息并克服奇异性和类别分离问题的线性特征提取算法；面向流形分布数据，研究基于多核学习和分类准则的图构造方法，以刻画数据的局部近邻关系和鉴别信息，研究图上的大间隔特征提取模型和快速算法，以提取数据的非线性特征并与模式分类建立直接联系；研究能融大间隔原理和稀疏性于一体的特征提取算法，以利于消除冗余信息，增强特征鉴别能力和可解释性；从特征生成、姿态子空间学习和回归估计等角度，研究对身份、光照、遮挡等因素鲁棒的驾驶员头部姿态估计模型和算法。本项目研究将推动特征提取方法的发展与应用，具有重要的理论意义和实际应用价值。

中文关键词： 特征提取；支持向量机；大间隔原理；优化算法；智能交通

英文摘要： Discovering most discriminatory features from data is a fundamental problem in pattern recognition. In order to deal with the high-dimensional small-sample-size data and manifold-distributed data, the project takes the advantage of frontier approaches in optimization, integrates sparse learning theory and aims to research large margin principle-based feature extraction model and algorithm and their applications to driver head pose estimation. With respect to high-dimensional small-sample-size data, from the viewpoint of the decomposition of high-dimensional space, the building of large margin model as well as parameters optimization, the project researches linear feature extraction methods which can make full use of all the discriminative information and overcome singularity and class separation problems. With respect to manifold-distributed data, the project researches graph construction based on multiple kernel learning and classification criterion, so as to describe the neighborhood relations and discriminative information. Building on the constructed graph, it further researches large margin nonlinear feature extraction model which is relevant to the subsequent pattern classification. The project researches feature extraction methods by combing large margin principle and sparsity, which can delete redundant

英文关键词： Feature extraction；support vector machine；large margin principle；optimization algorithms；intelligent transportation