基于多级动态模型的2维人体姿态估计  

2D Human Pose Estimation Using Multi-level Dynamic Model

作　　者：马淼[1] 李贻斌[1] 

机构地区：[1]山东大学控制科学与工程学院,山东济南250061

出　　处：《机器人》2016年第5期578-587,共10页Robot

基　　金：国家863计划(2015AA042201);国家自然科学基金(61233014)

摘　　要：提出一种利用多级动态模型来估计单目视频中的人体姿态的方法.首先,构建了一种多级动态人体姿态模型,该模型将人体姿态视为各部位姿态的铰接组合,用部位姿态的最优估计来逼近整体姿态的最优估计,从而解决了整体姿态估计带来的歧义性问题.其次,提出了一种通过构建虚拟姿态来计算视频相邻帧之间姿态一致性的算法,该算法能够有效利用视频中表观特征及运动特征的连续性,从而提高姿态估计精度.此外,使用粒子群优化算法用较小的姿态样本优化出最优部位姿态,并将最优部位姿态重组为最优的人体姿态.通过实验验证了所提方法的有效性,并与几种前沿方法进行了比较.实验结果表明,本文方法有效提高了人体姿态估计的准确度.A human pose estimation algorithm with a multi-level dynamic model for monocular videos is presented. Firstly, a multi-level dynamic model of human pose is constructed to decompose human entire pose into articulated pose parts,and approach optimal human pose candidates by optimizing pose parts candidates. This model solves the ambiguity problem caused by the entire pose estimation method. Secondly, an algorithm for calculating the pose consistency between the adjacent video frames is proposed by constructing virtual poses. This method can make use of the continuity of appearance features and motion features between the adjacent frames to improve the estimation accuracy. Thirdly, particle swarm optimization method is utilized to search for the best pose parts candidates with a small amount of candidates, and then the achieved pose parts are recomposed into the optimal human entire poses. The efficiency of the proposed method is tested and experimentally compared with several related state-of-the-art methods on challenging video sequences, which shows significant improvements.

关 键 词：人体姿态估计 多级动态模型 视频理解 人体行为理解 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]