基于区域分割和蒙特卡洛采样的静态图片人体姿态估计

Human pose estimation in static images based on region segmentation and Monte Carlo sampling

静态图片中的人体姿态估计是近年来图像分析领域的重要问题之一.由于静态图片中可利用的信息较少,且存在多关节引起的形状畸变、着装变化、背景干扰及遮挡等难点,使得这一问题具有很大挑战性.考虑到现有算法的不足,提出了一种基于区域分割和置信传播蒙特卡洛采样的人体姿态估计算法,将前景区域分割加入到姿态估计中,在概率图模型中引入非树状的约束,采用蒙特卡洛采样来进行概率推理.实验表明该算法比经典的算法在公共数据集上给出了更加精确的估计结果,同时运行时间也减少了25%.

Human pose estimation in static images is one of the important issues in the field of image analysis in recent years. The main difficulties are that there＇ s less available information in static image, besides, there＇ re figure distortion due to multiple joints, change in clothes, background disturbance and shading, etc. , which make the problem challenging. Aiming at the deficiency of existing algorithm, a new algorithm was proposed for human pose estimation in static images based on region segmentation, belief propagation and Monte-Carlo sampling, in which foreground region segmentation was incorporated into the pose estimation, non-tree constraints were introduced in the probabilistic graphical model, and Monte-Carlo sampling was utilized to carry out probabilistic inference. Experimental results demonstrate that the proposed algorithm performs better on a common database compared with classical algorithm, producing a more precise estimate result and reducing 25% running time.