基于张量与姿态回归网络的多视角多人姿态估计

Multi-person Pose Estimation Based on Tensor and Pose Regression Network

本文提出一种基于张量与姿态回归网络的多人姿态估计方法。所提出的方法与以往基于噪声和不完整姿态的交叉视觉估计法有所不同,所提出多视角二维图像像素处理可以直接在三维空间上处理数据,从而避免了每个视角姿态的错误判定。本文方法包括张量网络和姿态回归网络,两个网络的组合可以更精确地估计出多视角相应的三维姿态。本文方法对于解决姿态的遮挡估计和歧义问题具有很强的稳定性,经过一些常用实验数据集测试,其比前期的一些方法具有更强性能。

A new multi-person pose estimation method based on a tensor and pose regression network is proposed in this ar⁃ticle.The proposed method is different from the previous cross-visual estimation methods which are based on noise and incomplete pose.The multi-view 2D image pixel processing in the proposed method can directly process the data in the 3D space,thus the wrong determination of each view can be avoided.The proposed method is including tensor network and pose regression network,and the combination of the two networks can more accurately estimate the corresponding 3D pose from multi-view images.The pro⁃posed method has strong stability to overcome occlusion and ambiguity for the estimation of 3D poses.After the testing by some re⁃leased datasets,the experimental results are verifying that the proposed method has the better performance than some previous methods.