基于头肩HOG特征的快速行人检测

Fast Human Detection Based on HOG Feature of Head-shoulder Model

提出了一种基于头肩模型的快速人体检测算法,在当前方向梯度直方图与支持向量机相结合的人体识别方法的基础上,以头肩模型代替整个人体进行人体识别,并且采用混合高斯目标提取技术减小搜索范围,通过降低方向梯度直方图的维数提高识别速率。首先结合混合高斯目标提取技术与Sobel边缘检测技术获取运动目标轮廓,并计算头肩模型范围。计算头肩模型的HOG描述子并通过SVM分类器进行分类。最后,对分类为非人体的目标进行二次识别,克服混叠等因素造成的错判。实验结果表明,该算法在识别率和识别速率上都有所提高,并且在骑车等特殊行人检测中也有很好的识别效果。

In the paper, a novel method for fast human detection is proposed. This method develops from the current mainstream human detection method, which is known as the HOG ＋ SVM method. But differently, HOG （ Histogram of Oriented Gradients） features of the whole body is replaced by that of the head-shoulders in the paper. The method contains four parts. Firstly, the head-shoulder model is extracted from the object image that segregated by Gaussian Mixture Model. Sobol is used in the process to find the contour of human body,and binary image got by Gaussian Mixture Model is used to a- mend the contour auxiliarily. Then,the head-shoulder model is extracted according to the normal proportion of human beings. Secondly, the HOG fea- tures of the head-shoulder model are calculated. Thirdly,the HOG feature of human head-shoulder model is put into the SVM to find out whether the input is the person or the nonperson. When the human body is occluded in some cases, the fourth step which called reclassification is need to be carried out. The experimental results show that the human body detection rate and speed are significantly improved compared with those of traditional method which deal with the whole human body.