Abstract: Aiming at real-time deficiency based on histograms of oriented gradient in pedestrian detection, this paper proposes a new method which is based on multi-detector fusion. It uses Adaboost classifier with Haar feature to detect head roughly, however, it exists some undetected cases in the process because of the pedestrian poses or dimension problems in the images. It adopts Canny operator to get image contour and obtains image contour based on color feature, then it extracts the possible detective area in the image by using ellipse fitting. According to the previous rough detection results and through suitably changing the scale of candidate region, it gets PHOG feature and judges by using linear SVM. It is approved that the new method is efficient and available in the test of INRIA sample database.

Key words: pedestrian detection, Pyramid of Histograms of Oriented Gradient（PHOG） feature, color feature, Adaboost, linear Support Vector Machine（SVM）

摘要： 针对基于梯度方向直方图的行人检测尚存在实时性不足的问题，提出了一种多检测器融合的行人检测方法。利用Haar型特征的Adaboost进行头部粗检，由于图像行人姿态或尺度的原因会导致这一过程出现漏检；采用Canny算子获取图像轮廓，并根据颜色信息获取图像轮廓，通过椭圆拟合提取图像中可能检测区域；根据前面粗检的结果，对候选区域合适变换尺度，提取PHOG特征，并采用线性SVM对其进行判决。在INRIA样本库上的测试结果表明该方法是有效可行的。

关键词: 行人检测, 金字塔式梯度方向直方图（PHOG）特征, 颜色特征, Adaboost, 线性支持向量机