一种基于HOG的多清晰度模板行人检测方法

Multi-Definition Template of HOG for Pedestrian Detection

行人检测是计算机视觉领域一个非常活跃的研究方向,为消除HOG特征提取过程中混叠现象对行人检测的影响,在特征提取过程中采用三线性插值技术;为解决待检测图像上采样过程会造成原始图像模糊而影响提取到的HOG特征不能完整描述行人的问题,提出一种基于多清晰度模板的分类器训练方法,首先对测试样本进行模糊处理,然后再提取HOG特征进行分类器训练用于行人检测。在INRIA行人检测库和自建的行人库进行了广泛的实验,仿真结果表明在窗口误检率(FPPW)为10-5量级时,将漏检率降低到了15%;在640*480的图像分辨率下检测时间控制在0.5s内。

Pedestrian detection is a very active field in computer vision. In order to reduce the influence of overlap between adjacent blocks,a tri-linear interpolation is applied in the steps of extracting HOG feature. To overcome the blur brought by the up-sample of image and poor description of HOG,a multi-definition SVM classifier is proposed.First,the test image sample is blurred,then HOG feature is extracted to train classifier and detect pedestrian. Comprehensive experiments on INRIA Pedestrian database and self-built database show that the miss rate of detection is decreased to 15% at the false positive per window（ FPPW） of 10-5,and the average detection time is about 0. 5s on the image resolution of 640＊ 480.