联合梯度直方图和局部二值模式特征的人体检测

HOG-LBP pedestrian detection

针对采用单一梯度方向直方图(HOG)特征进行人体检测时易受竖直梯度分量干扰的缺点,提出了将分块局部二值模式(LBP)特征加入HOG特征的方法。首先,将检测窗口划分为大小为16×16的不重叠块,以块为单位统计LBP特征直方图,并通过大量实验获得了LBP算子的最佳参数;然后用优化过的插值方式计算HOG特征,将两者组成联合直方图。最后,用线性支持向量机(SVM)通过Bootstrapping的方式训练,得到判别模型。在INRIA人体库上的测试表明,检出率在误检率(FPPW)为10-4时由原始的89%提高到95%,单窗口检测速度由0.625ms提高到0.533ms。本文将纹理特征加入原始描述轮廓的HOG特征中,排除了部分梯度干扰信息造成的误检,提高了检出率。

This paper proposed a method to concatenate a cell-structured Local Binary Pattern（LBP） feature into Histogram of Gradients（HOG） to solve the problem that HOG was vulnerable to the interference of ver- tical background gradient information in pedestrian detection. Firstly, the detection window was divided into 16 × 16 non-overlapping blocks, then the LBP histogram of each block was calculated and his parameters were obtained by extensive experiments. Afterwards, the HOG was computed by the optimized interpolation meth- od, and it was combined with LBP histogram to constitute a joint histogram. Finally, a discriminative model was trained by Bootstrapped linear Support Vector Machine（SVM）. Based on the test of the INRIA pedestri- an dataset, it is shown that the detection rate has been increased from 89% of the HOG feature to 95% when False Positive Per Window（FPPW） is 10-4 ,and the detection speed has been raised from 0. 625 to 0. 533 ms per window. It is concluded that the proposed method in this paper eliminates the false detection caused by the interference of gradient information and improves the detection rate by describing both contour and texture in- formation.