3D Human Pose Estimation Based On Deep Adversarial Learning

3D human pose estimation is to accurately locate the key point of a human body in the image or video,and abstract a 3D human bone shape for subsequent action analysis and behavior recognition.It is the basis of many advanced computer vision tasks and can be widely used in video surveillance,gait analysis,motion capture and human-computer interaction.With the rapid development of deep neural network,the method of 3D human pose estimation is being innovated day by day.The two-stage 3D human body pose estimation method based on deep learning firstly inputs the image into the pre-trained 2D pose extraction network to obtain the coordinates of 2D human pose.Then a neural network establishes the mapping from 2D pose to 3D pose point.This kind of two-stage method shows good results in practical application.However,in the transformation from 2D pose to 3D pose,most of the existing methods carry out supervised iterative training through a large number of labeled data.On the one hand,it ignores the effective information in the distribution of the data,on the other hand,it causes the over-fitting phenomenon on a single data set.In practical application,the scarcity of 3D human posture labeling data also restricts the further application of supervised methods,which can not meet the training of diversified 3D human body estimation network in natural scenes.In recent years,with the rapid development of Generative Adversarial Networks technology,the information of data distribution can be extracted effectively through the adversarial learning of the network.Therefore,in this paper,under the framework of deep adversarial learning,the problem of 3D human pose estimation in different training scenes is studied by using Generative Adversarial Networks(GAN).The specific contents are as follows:1.Most of the existing supervised 3D pose estimation methods do not take into account the distribution characteristics of pose data when reconstructing 3D human pose.Taking this as a starting point,this paper proposes an pose projection generative adversarial networks for supervised 3D attitude estimation.A special projection discriminator is constructed to provide two new constraint information for the supervised model,so as to improve the accuracy of 3D human pose estimation.In this paper,Wasserstein GAN with Gradient Penalty(WGAN-GP)and Least Squares GAN(LSGAN)are used to verify the effectiveness of the model.Qualitative and quantitative analysis shows that the accuracy of the supervised model can be effectively improved by introducing new constraints into the projection discriminator.2.Effectively learning the distribution information of pose data can further reduce the dependence of the training process on the real 3D pose label data.In this paper,based on the generative adversarial network WGAN-GP,using the structural characteristics of pose data in the projection transformation,a weak supervised cycle projection network and an unsupervised dual direction projection network are designed for 3D pose reconstruction.The cycle projection network does not need a clear corresponding relationship between 3D pose and 2D pose attitude data in the training process,and the bi-directional projection transformation restricts the reconstruction process of 3D pose from two aspects,fully excavates the physical characteristics of the pose data,and can complete the training of the model without relying on 3D pose data.the experiment achieves competitive reconstruction results on both MPII and Human3.6M data sets.