Research On Frontal Human Body Segmentation Algorithms In Static Images

With the improvement of living standards and deepening of social informationization, digital images more and more become an important part of acquiring information in everyday life. As an essential component of image processing, as well as indispensable step of image pre-processing, image segmentation has important research value for the subsequent target detection, image retrieval and image understanding.Human segmentation in video sequences is common nowadays. It is mainly used in monitoring systems of public places. Additionally, most of these methods there have two common drawbacks: the moving target detected may not be the human body; and static human body cannot be detected. This paper focuses on the study of detection and segmentation of human bodies in single static images. It is of great significance for the research and application of content-based image retrieval, human-object detection and classification and image understanding.Among algorithms of human body detection, this paper describes several existing methods of human body detection analyses their reliability, and on the basis of it propose an algorithm of human body detection and segmentation based on face detection and Normalized Cuts. Colors, shapes and gestures of human bodies vary greatly, which brings in numerous difficulties to the human body detection, so this paper adopts a two-step approach:1) To detect upper body, including the head, torso and arms. Under the premise the position and size of the face is known, we first use template matching to find the approximate region of upper body with an optimal angle, then use graph segmentation to accurately separate out the torso region; meanwhile also detecting the skin-color region. Finally we combine the torso region, the face region and the skin-color region together to get a complete upper body region;2) Based on segmentation result of the upper body, we use the main-color estimation to find the approximate location of the hip, and then use graph segmentation again to separate out the accurate and complete leg region.The advantage is the ability to overcome the vexation brought to human body detection by different colors, shapes and relative positions of lower bodies. Experimental results show that the method proposed in this paper has good effect on the segmentation of frontal humans and is robust.