基于3D heatmap的人体三维姿态估计方法

Human Pose Estimation Method Based on 3D Heatmap

针对肢体间自遮挡和物体遮挡造成的关节位置信息缺失问题,提出了一种基于3D heatmap的三维姿态估计方法。首先,采用卷积神经网络对人体二维关节点热图heatmap进行提取,并根据各摄像头间的外参数构建人体各关节点的初始3D heatmap;然后,利用人体扫描与建模获得的人体模型先验信息,采用期望最大化算法,迭代优化获得最符合人体模型的关节点在空间的位置分布。最后利用Human3.6M数据集进行验证。结果表明:在部分视图存在遮挡的情况下,可获得高精度的三维关节点位置。

To deal with the lack of joint position information caused by the self-occlusion between the limbs and the occlusion from objects,a 3D heatmap-based 3D pose estimation method was proposed.First,a convolutional neural network was used to extract the heatmap of the human body,and the initial 3D heatmap of the human body was constructed according to the extrinsic parameters of the multi-view cameras.then,using the prior information of the human model obtained from human body scanning and modeling,the expectation maximization algorithm was used to iteratively optimize the spatial position distribution of the joint points that most consistent with the human model.Experiments on the Human3.6M data set showed that high-precision three-dimensional joint point positions could be obtained in the presence of occlusions.