中文摘要：

针对传统的梯度方向直方图（HOO）行人检测方法计算复杂、实时性较差的问题，提出了一种改进的HOG行人检测方法。首先，利用高斯混合模型背景建模，去除大部分背景图像，减少滑动窗口扫描区域，以提高检测速度。同时．选择头肩特征作为行人检测依据，计算多尺度HOG特征，减少计算量，降低因姿态变化遮挡等引起的误检测率。通过行人头肩特征图像库的实验证明，该方法能有效提高检测速度，并得到较高的检测精度。

英文摘要：

Aiming at the complex process and the poor real-time of the traditional HOG（histograms of oriented gradient） pedestrian detection method, this paper presented an improved one. First,the mixture Gaussian model was used to build a background model, the majority of the background image removed, the scan regions of the sliding windows reduced and the detection speed improved. Then the multilevel head-shoulder HOG feature was selected as the measure to detect and compute the pedestrian multilevel HOG, reducing the computation and also lowing the false detection rate due to the attitude change or part body sheltered. Finally, the experimental results of our own head- shoulder pedestrian database show that this improved method can improve the detection speed effectively and get a higher detection accuracy.