基于LPP的视频图像头部姿态估计的方法研究

【作者】 陈书明；

【导师】 陈锻生；

【作者基本信息】 华侨大学 ， 计算机应用技术， 2011， 硕士

【摘要】 头部姿态估计在注意力检测、行为检测和人脸识别上有着重要的研究意义。局部保持投影算法为人们处理非线性降维问题提供了一种新的思路。作为一种线性降维方法,在头部姿态估计领域具有很大的应用空间。对于头部姿态估计问题,本文研究主要围绕有监督的局部保持投影和异常值的度量方法进行展开。鉴于无监督局部保持投影的流形学习算法对头部姿态估计的高误差性和对噪音的敏感性问题,本文设计正弦偏置距离方法和融入带权值主成分分析方法来改进局部保持投影算法。首先对训练的头部姿态加以姿态标注,并获得各个头部姿态间的正弦偏置距离;然后对所有头部姿态样本点进行异常值的度量,训练出较好的线性映射矩阵,再采用改进后的局部保持投影算法对图像进行降维处理;最后采用支持向量机分类器进行头部姿态估计。用改进的局部保持投影方法进行了头部姿态估计实验。由于融入了正弦偏置距离和带权值主成分分析方法,不仅有效地消除人的身份的影响,还很好地削弱光照变化、表情变化、噪音等因素的影响。并且大量实验也表明:改进后的局部保持投影算法同改进前局部保持投影算法、局部嵌入分析算法相比,无论在静态的头部姿态数据库中,还是在动态的视频流中,头部姿态估计都取得较好的效果。更多还原

【Abstract】 Head pose estimation is an important research issue in attention test, behavioral detection, and face recognition. Locality Preserving Projection method provides a fresh idea for us to do nonlinear dimensionality. As a linear dimension reduction tool, Locality Preserving Projection method has great potential to be applied in head pose estimation.This research concentrates on Supervised Locality Preserving Projection and Outlier Measure method to start. Aiming at the problems of the high head pose estimation error and the noise sensitivity of unsupervised LPP, Sinusoidal offset distance method and weighted PCA method was used to estimate the head pose. By this algorithm, the head poses of the training samples are firstly labeled, and all the biased distances between the head poses are calculated. And then the outliers of the head samples are detected, so as to train a better linear mapping matrix, then using the Improved LPP to reduce the dimensions of the image; the SVM classifier are finally used to estimate the head pose.In this paper, the Improved LPP method is used to estimate the head pose. Sinusoidal offset distance method and weighted PCA method not only effectively eliminate the impact of the identity, but also effectively reduce the impact of illumination, facial expression changes and noise. The head pose estimation experiments show that the improved LPP achieves the better results than the traditional LPP and LEA algorithms in both static head pose database and dynamic video stream.更多还原