The Research Of Human Body Posture Estimation Method Using Video Images

The estimation of human body posture from a video image is a new technology, which recovering human body 3D posture by analyzing and disposing underlying characteristics in computers. As the development of computer vision technology and image processing techniques, in recent years, the posture estimation of a human body form an image has been put on a focal position in research.This article is based on the key frames, which provided by fellow researchers. In order to solve the shortcomings of existing estimated method at low speed and the partial adjustment cannot be done on estimated posture. A method of human posture estimation based on partial silhouette is proposed in this issue. It mainly about works as followed:Firstly,the human body is a whole which is made up by a head, four limbs and trunk . When we analyze human body, have to consider the influence of every part to the whole. Therefore, a method of decomposing human body silhouette is put forward. In this method, body silhouette is primarily divided into segmentations depends on the negative curvature extreme points, under the limit of the ratio between arc length and secant line. Then, make use of the positive curvature extreme points which correspond to each silhouette segment, to combine fine silhouette segments. So we can obtain the silhouette, which can represent limbs of human body. Finally, we can obtain two kinds of silhouettes, main trunk segments and limbs segments.Secondly, in order to improve the speed of posture estimation, we choose silhouette segments to match in the matching stage. At first, it gets the correspondence of silhouette segments by using the extreme point of each segment. Then, we use the character of limbs segments to reduce calculation in the matching corresponding segments. At last, the similarity of the whole silhouette is evaluated by means of the statistical analysis of the similarity of every part. We select the most similar sample from silhouette database, and using its corresponding 3D data as the initial estimation.Thirdly, we get the part of initial posture estimation which needs for pose adjustment by using the similarity of the silhouette segments, use the affine transformation of silhouette points to transplant the skeleton point, restore the three-dimensional coordinates of skeleton points by camera affine projection model. Then, we get more real posture estimated results.Experimental results show that the method can effectively improve the speed of the body posture estimation from a video image and make the final estimation results more close to real human body posture in the video.