基于RGB-D图像弱监督学习的3D人体姿态估计

3D human pose estimation based on RGB-D image weakly-supervised learning

针对于深度图数据缺乏大量的3D标签、泛化能力差的问题,结合现有的弱监督网络结构,提出一种基于RGB-D图像的弱监督模型实现3D人体姿态估计的方法,整体呈现两级级联结构。首先通过使用预处理后的RGB-D数据作为2D姿态估计模块的输入,提取出人体关节热图;然后将热图进行积分回归生成对应的关节点坐标;最后将生成的关节点作为改进型深度回归模块的输入完成姿态估计。通过在公开数据集Human 3.6M和ITOP上进行验证,实验结果表明:本文提出的弱监督网络模型在参数量上减少了20.9%,训练时间上降低了37.9%。提出的模型能同时适用于深度图和彩色图,且具有较强的鲁棒性。

Aiming at the problem of the lack of a large number of 3D labels in the depth map data and the low generalization ability,combined with the existing weakly supervised network structure,this paper proposes a method of 3D human pose estimation based on the weakly supervised model of RGB-D images,whole network presents a two-level cascade structure.First,By using the preprocessed RGB-D data as the input of the 2D pose estimation module,the human joint heat map is extracted;then integrate the heat map to generate the corresponding joint point coordinates;and finally use the generated joint point as the improved depth The input of the regression module completes the pose estimation.Through verification on the public datasets Human 3.6 M and ITOP,the experimental results show that the proposed weakly supervised network model reduces the parameter amount by 20.9%and the training time by 37.9%.The proposed model can be applied simultaneously used in depth maps and color maps,and has strong robustness.