为提高智能视频监控中行人统计的实时性,提出了一种基于人头颜色空间和轮廓特征的行人检测方法.该方法首先根据人脸肤色、发色在YCbCr和RGB颜色空间的聚类情况,建立人头颜色模型,分割人头候选区域,并针对行人运动的特征,采用多帧差法提取运动信息,剔除背景噪声,修正候选区域的精度;然后根据改进的Canny算子提取候选区域的轮廓,融合形态学对边缘进行修正,提取候选区域轮廓信息;最后根据人头轮廓的几何特征,剔除"伪候选"区域,并进行连通域信息标记,检测人头图像,从而对行人进行检测和统计信息.结果表明,该方法能快速有效地检测出人头,在动态场景下的行人检测取得了较好的效果.%An approach of pedestrian detection based on the head color space and outline information is proposed to meet the requirement of reducing people counting time in real-time video monitoring. A head color model is established according to head feature in YCbCr and RGB color space,which obtain head candidate areas. Multi-frame difference method is adopted to eliminate background noise and correct candidate regions based on the characteristic of pedestrian movement. The outlines of candidate regions are extracted based on improved Canny operator and morphology. Fade candidate regions are removed according to the geometrical characteristics of head contours, the connected components are marked to detect pedestrian information. Experimental results show that this algorithm is fast and effective, which could meet the real-time needs of intelligent video surveillance, the better pedestrian detection results are achieved in dynamic traffic scene.
展开▼
