利用卷积曲面模型的人体运动姿势估计

The Human Pose Estimation from Monocular Image

基于通过单目图像进行人体运动姿势估计,是计算机视觉中非常困难的问题,提出了一种利用单目图像进行人体运动姿势估计的框架.首先建立了一种新颖的多项式密度卷积曲面人体模型,该模型由连接体线骨架与某种卷积核卷积而成,整个曲面是一张连续曲面,调节多项式和半径参数可以让人体任意变形;给出了正交投影下卷积曲面和曲线对应的概念和相关定理,为二维的图像轮廓和三维人体模型之间建立对应关系;根据正交投影下卷积曲面与曲线对应定理,用卷积曲线去逼近人体的图像轮廓曲线,从而估计和恢复出三维人体运动姿态.实验证明,该方法在没有遮挡的情况下,可以很好地恢复人的三维运动.

A new human pose estimation framework was presented to recover human pose from monocular image. A human model using polynomial convolution surface is presented, which is modeled by articulated skeletons convolved with some kernel function. The convolution surface and curve correspondence theorem under orthogonal projection is presented. The human contour is approximated with a convolution curve and 3D human pose can be estimated using a nonlinear optimization algorithm. The experiment shows the method is efficient without serious occlusion.