基于分步降维HOG-LBP特征的行人头部分类算法

Pedestrian head classification algorithm with two-step dimension reduction HOG- LBP feature

传统的基于PCA-HOG特征的行人头部分类算法存在降维后的子空间鉴别性不足的问题.为此,提出一种基于分步降维HOG-LBP特征的行人头部分类算法.首先,利用样本类别标签构建2类样本的HOG特征集合,在这2类特征集合中分别执行PCA降维,然后将所得的特征与LBP纹理特征进行级联得到最终的头部描述算子,最后通过训练SVM分类器对实际样本集进行分类.实验结果表明,与传统PCA降维方法相比,该方法可有效提高行人头部的分类性能.

Traditional pedestrian head classification algorithm based on PCA-HOG feature has the problem of degradation of the discrimination in the subspace.In order to handle this problem,the pedestrian head classification is completed based on the proposed two-step dimension reduction HOG- LBP feature.Firstly,two category of HOG sample set are obtained according to the sample labels.The PCA algorithm is carried out on each sample set step by step.Then the LBP texture features are combined with the dimension reduced HOG feature to form the final head descriptor.Lastly,experiments were performed by SVM classifier on practical test samples,and the experimental results show that,comparing with the traditional PCA algorithm,the presented HOG-LBP features can effectively improve the classification performance of pedestrian head.