结合轮廓粗筛和HOG细分的快速行人检测方法

摘要

For solving the problems of high false alarm rate and low efficiency of existing pedestrian detection methods,a fast pedestrian detection method is proposed.First,this paper executes contour rough selection of the first layer according to geometric features of the contours,and executes the one of the second layer according to the invariant moments of the contours,to remove false-targets and reduce false alarm rate.On this basis,this paper crops the suspicious image patches and only extracts histogram of oriented gradients(HOG) features on these image patches,and then uses linear support vector machines to classify these features,for further reducing the false alarm rate.Meanwhile,the efficiency is improved because the number of HOG feature extraction is sharply reduced.Simulations train and verify features on INRIA and Caltech dataset,respectively.Results show that the new method has low false alarm rate and high efficiency.%为解决现有行人检测方法虚警率高、运算效率低的问题,提出一种快速行人检测方法.首先,依靠轮廓的几何特征进行第一层粗筛选,依据轮廓的不变矩特征进行第二层粗筛选,剔除干扰目标,降低虚警率.在此基础上,裁剪图像上的可疑图像块,仅在可疑图像块上提取HOG特征,并结合线性支持向量机进行特征分类,进一步降低虚警率.同时由于大幅减少了提取HOG特征的数量,从而提高了运算效率.仿真实验在INRIA数据集上进行训练,在Caltech数据集上进行验证.结果表明,该方法的行人检测虚警率低,运算效率高.