摘要
基于混合部件模型的人体姿态估计方法忽视了人体结构的对称位姿约束关系,从而导致对称部件容易被重复检测、人体姿态估计准确率较低,为此,提出一种基于位姿约束与轨迹寻优的姿态估计新方法。首先估计人体单部件和对称部件在单帧图像中的多个合理位置,利用对称部件之间的位姿约束关系构建标识部件。然后根据单部件和标识部件各自的目标优化函数,通过动态规划算法反复迭代获得初始轨迹候选集,再结合轨迹的全局特征剔除检测得分较低的运动轨迹。最后引入树形合约模型,联系时空上下文信息,准确求解出视频序列光滑且兼容的最优轨迹。在N-best、Outdoor Pose和Scene数据集中的实验结果表明,对于存在背景复杂、运动模糊、部件遮挡等问题的视频序列中,该方法平均姿态估计准确率达87%以上,有效减少了对称部件的误判,提高了视频中人体姿态估计的准确率。
Because of ignoring the configuration constraints between symmetric body parts, the human pose estimation methods based on mixtures of parts may lead to a repetitive detection of symmetrical body parts and a low pose estimation accuracy. Therefore, a kind of new pose estimation method on the basis of pose constraint and trajectory optimization was put forward. Firstly, numerous reasonable locations of single part and symmetric parts of human in single-frame image should be estimated, and identification part should be constructed by utilizing pose constraint relationship among symmetric parts. Then initial trajectory candidates set shall be gained through repeated iteration of dynamic pro- gramming algorithm according to respective target optimization function of single part and identifica- tion part. Movement trajectory with relatively low detection score was removed by combining with global feature of trajectory. Finally, smooth and compatible optimal trajectory of video sequence wascorrectly solved by introducing tree-based contract model and combining with contextual spatio-tempo- ral information. Experimental result in N-best, Outdoor Pose and Scene dataset shows that in video sequence with complex background, blur movement and part blocking problems, average pose estima- tion accuracy of proposed method is greater than 87%, which reduces erroneous judgment of symmet- ric parts effectively and improves human pose estimation accuracy in video.
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2017年第4期1060-1069,共10页
Optics and Precision Engineering
基金
国家自然科学基金(41301448)
江苏省重点研发计划(BE2016071)
关键词
人体姿态估计
混合部件模型
位姿约束
最优轨迹
Human pose estimation
mixed parts model
pose constraint
optimal trajectory
作者简介
李庆武（1964-），男，河南新乡人，博士、教授、博士生导师，1985年于郑州大学获得学士学位，1990年于西安电子科技大学获得硕士学位，2010年于河海大学获得博士学位，主要研究方向为智能感知与图像处理。E—mail：liqw@hhuc．edu．cn
席淑雅（1993-），女，河南商丘人，硕士研究生，2015年于河海大学获得学士学位，主要研究方向为数字图像处理。E—mail：xishuya@hhu．edu．cn
