基于人体模型和超像素的黏连人群分割方法

Overlapping crowd segmentation based on human model and superpixels

针对监控视频中拥挤人群的分割问题,提出一种对黏连人群进行分割的新方法.该方法利用投影法和Hough圆检测方法对目标的头部进行检测,利用卡尔曼滤波跟踪算法实现遮挡情况下的人群头部粗定位;采用人体模型实现行人的粗分割;对目标前景进行超像素分割,基于相邻像素块之间的颜色相似程度和与人体模型的形状匹配程度构建一个加权图模型,通过求解最优路径的方法得到黏连目标的最优分割边界.实验表明:该方法能有效解决黏连人群的分割问题,且能够精确地提取出完整的人体轮廓.

Aiming at the problem of crowd segmentation for the video surveillance, a new method of segmenting individual humans in overlapping situations was proposed. In this method, vertical projection histograms and Hough circle transformation were used to detect human＇s heads, Kalman filter was used to help locate the positions of the heads when occlusion occurs, human shape models was used to segment pedestrian roughly, the foreground area was segmented into several superpixels, and the best segmentation boundary of the overlapping crowd is defined by the optimal path with a weighted graph model based on the dissimilarity between adjacent regions＇ color and the degree of mismatching of a human model. Experimental results show that this method can segment the overlapping crowd effectively and extract the human body boundary precisely.