Bagging-SVM集成分类器估计头部姿态方法

Head Pose Estimation Method of Bagging-SVM Integrated Classifier

针对现有常用分类器性能不能满足头部姿态估计对准确率的要求,以及光照变化影响头部姿态估计准确率的问题,提出了一种基于Bagging-SVM集成分类器的头部姿态估计方法。首先,通过图片预处理和Adaboost检测人脸区域算法减少背景、光照等干扰因素对于头部姿态特征提取的影响。其次,采用融合方向梯度直方图(HOG)特征和局部二值模式(LBP)特征分别对人脸的轮廓特征和纹理特征进行提取。然后,通过主成分分析(PCA)对融合的头部姿态特征进行特征选择,抽取其主元特征分量供分类器进行训练。最后,通过Bagging方法构建多个训练数据集,并采用支持向量机(SVM)对每个数据集进行训练,产生多个弱分类器,多个弱分类器投票决定测试样本所属类别。将该算法在Pointing’04数据集、CAS-PEAL-R1数据集和自建数据集上进行验证实验,实验结果表明提出的算法相比线性判别分类器(LDA)、朴素贝叶斯分类器(NB)等常用分类算法具有更高的分类准确率,对光照的变化具有较好的鲁棒性。

Aiming at the problem that the performance of existing common classifiers can’t meet the requirements of head pose estimation accuracy and the influence of illumination changes on head pose estimation accuracy,a method about head pose estimation based on Bagging-SVM integrated classifier is proposed.Firstly,the image preprocessing and the detection of face area based on Adaboost algorithm are used to reduce the influence of background,illumination and other interference factors on head pose feature extraction.Secondly,the histogram of orientation gradient(HOG)and the local binary pattern(LBP)are fused to extract the contour features and texture features of the face respectively.Then,principal component analysis(PCA)is introduced to select the fused head pose features,and the principal component features are extracted for the classifier to train.Finally,several training datasets are obtained by Bagging,each dataset is trained by support vector machine and several weak classifiers are produced.The categories of the test samples are voted by the plurality of weak classifiers.Experimental results on the Pointing'04 dataset,CAS-PEAL-R1 dataset and self-built dataset show that the proposed algorithm has higher classification accuracy than the commonly used classification methods such as linear discriminant analysis(LDA)and naive Bayesian(NB)classifier,and has better robustness to illumination changes.