基于颜色和变形模板的实时人体检测

Color and Deformable Templates Based Real-time Pedestrian Detection

针对目前人体检测算法存在不能检测多角度人体目标以及实时性差等问题,提出了一种在静态图像中实时检测任意角度人体目标的算法。该算法分别利用目标颜色和轮廓两类特征构造两种检测器。颜色检测器首先进行基于面部肤色和头部发色的彩色分割,然后引入积分图像算法快速提取分割后的图像头部目标区域。轮廓检测器利用头肩轮廓形状的稳健性,用参数化变形模板对头肩轮廓建模,该模板由两个存在几何尺度和位置约束的椭圆构成,再定义两个不同计算复杂度的模板匹配策略对人体头肩部分进行分级检测。最后利用上述两种检测器构建一个级联检测系统,级联检测结构大大提高了算法的速度,使算法可以对分辨率为352×288的图像做30 fps的实时检测,实验结果表明,该算法是切实有效的。

Most of the present pedestrian detecting algorithms can not deal with multi-view objects and can not perform in real time. Faced with these problems, this paper provided a real-time multi-view pedestrian detecting algorithm, which utilizes respectively the color and the contour features to construct two detectors. One is a color based detector, which, first, segments the original images based on skin color of face and hair color, and then locates the candidate head regions by using integral images, and the other is the contour detector, which employs the robust head-shoulder contour feature to construct a deformable template based on two ellipses with size and position constraints. Two template matching algorithms with different computing complexity were introduced to detect head-shoulder objects. Finally, a cascaded detecting system was generated by the two detectors mentioned above. The whole system can process images with 352 x 288 resolution at a speed of 30fps by using the cascaded structure. The algorithm proved to be effective.