为满足人流统计实时性的要求，提出一种人头检测算法。该方法对双目相机采集的图像通过运动目标检测分离出运动人员所在区域，对这些区域使用融合区域匹配和特征匹配的快速匹配方法，即利用视差的连续性只对强纹理点进行绝对误差累积(SAD)匹配，其他点只进行简单的视差验证，能够得到稠密的视差图，再由三角投影关系计算出深度图。由于双目立体成像得到的深度图中人员与场景的深度分布不同,可以采用深度分层的方法将存在人头信息的深度层提取出来，经过数学形态学预处理再利用边缘检测会得到许多候选轮廓，最终利用轮廓的几何特征来判断轮廓是否为人头。实验表明：该算法可以很好地适应复杂场景下的人头检测，精度高、速度快。

To meet the requirements of real-time statistics of the stream of people, a head detection algorithm is presented. The method isolates the region with moving persons by the moving target detection in the binocular camera images, a rapid matching method mixed region matching and feature matching is used to this region. Using the continuity of disparity, we just need to match the strong texture points by sum of absolute differences (SAD) method and simply verify the disparity to other points, so we can get the dense disparity map, and then calculate the depth map by the triangle projection relation. As the depth distributions between people and secne are different in the depth map generated by binocular stereo imaging, we can use the mathematical morphology operations to precondition, then get many candidate contours by edge detection. The characteristics of the head are used to determine whether the contour is the head. Experiments show that this algorithm can be well adapted to head detection in complex scenarios with high precision and speed.