基于改进时空特征的三维人体位姿估计方法  

Three-dimensional body poses prediction approach based on modified spatio-temporal feature

作　　者：谢立靖[1] 任胜兵[2] XIE Li-jing;REN Sheng-bing(College of Information,Hunan Vocational College of Commerce,Changsha 410205,China;College of Computer Science,Central South University,Changsha 410083,China)

机构地区：[1]湖南商务职业技术学院信息学院,湖南长沙410205 [2]中南大学计算机学院,湖南长沙410083

出　　处：《计算机工程与设计》2018年第11期3520-3525,共6页Computer Engineering and Design

基　　金：湖南省教育厅科学研究基金项目(14C0272)

摘　　要：针对传统三维人体位姿估计方法后期处理易受自我遮挡或镜像的影响,提出直接从边界盒的时空体估计中央帧的三维位姿的方法。选择连续帧中目标周围的边界盒,利用卷积神经网络进行运动补偿,使目标始终位于中央位置,形成修正时空体,从时空体中提取三维梯度方向直方图特征,作为修正时空特征,训练回归器估计中央帧的人体三维位姿。对Human3.6m、HumanEva和KTH多视角足球3种数据集进行实验,实验结果表明,与传统方法相比,提出方法通过综合利用外观信息和运动信息,三维人体位姿的估计准确性得到明显提高,泛化性能更强。To deal with the problem that the traditional method of 3D human body pose estimation is easy to be blocked by the self-occlusion and mirror image in a post-processing step,a method was proposed to estimate the pose of the central frame directly from the time and space of the bounding box.Bounding boxes around people in consecutive frames were selected.The convolutional neural network was used to compensate for the motion to make the people locate in the central location always to form spatio-temporal volumes.Three-dimensional histograms of oriented gradients were extracted as modified spatio-temporal feature.Regressor was trained to achieve body poses prediction.Through the experiments on the Human3.6m,HumanEva and KTH multiview football datasets,the results show that,compared with the traditional methods,the proposed method can get better body poses prediction effects and generalization ability,which combines appearance and motion cues.

关 键 词：人体位姿估计 运动补偿 卷积神经网络 核维纳滤波 深度学习网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]