中文摘要：

行人检测是目标识别领域的一大难题,针对行人检测存在特征维度高、检测耗时和精度低等问题,文章提出使用多尺度分块方式将样本图片在3个尺度下分别分割成5个区域,在每个区域中根据行人轮廓置信模板和梯度方向量化权值进行二次加权统计得到梯度直方图（histogram of oriented gradient,HOG）,并将其与Sobel边缘局部二元模式（Sobel edge local binary pattern,Sobel-LBP）算法相融合作为特征,然后采用线性支持向量机（support vector machine,SVM）分类方法学习得到行人检测分类器,最后使用滑动窗口法检测出行人。在MIT和INRIA库上的实验证明,该特征在学习和检测速度上都比HOG等方法有明显优势,能有效、准确、快速地检测行人。

英文摘要：

Pedestrian detection is a major difficulty in object recognition. To further improve the detection rate, decrease the large dimensions of features and solve the time-consuming problem in pedestrian detection, multi-scale blocks mode is integrated to extract features. Firstly, the sample image is divided into five regions in three scales, and the weighted histogram of oriented gradient（HOG） is obtained by corresponding confi- dence level according to the pedestrian outline incredible template and gradient quantitative weights in every scale, which is combined with Sobel edge local binary pattern（Sobel-LBP） as detecting feature. Afterwards, the linear support vector machine （SVM） classification is learned and used to detect pedestrian with sliding window method. Experimental results at the MIT and INRIA test set indicate that there are obvious advanta- ges in the learning and testing speed compared with HOG and several other methods. The proposed method is effective, accurate and rapid in pedestrian detection.