Research On 3D Human Pose Estimation Based On Generative Adversarial Networks

With the rapid development of artificial intelligence technology,the field of computer vision has been widely studied and applied.3D human pose estimation is a hot issue in the field of computer vision.Its task is to locate the joint positions of human bodies from images or videos and make the optimal connection.It is widely used in action recognition,human-machine Interaction,intelligent security,augmented reality and other fields,which has extremely high research significance and application value.In recent years,the generative adversarial network has made great progress in the field of image synthesis.The use of generative adversarial network can better extract human pose features and improve the performance of human pose estimation algorithms.Based on the generative adversarial network,this thesis conducts research on human pose estimation from both supervised and unsupervised perspectives.The details are as follows:Aiming at the problem that supervised 3D human pose estimation only pays attention to the accuracy of joint positions and ignores the rationality of the overall structure of the pose,this thesis proposes a supervised 3D human pose estimation algorithm based on generative adversarial networks.By combining the adversarial loss with the joint position loss,the network can predict more accurate 3D joint coordinates while learning more feasible human body structure.In order to learn the relationship between human joint points,this thesis introduces graph convolutional network,which can capture the local and global semantic information of human joints.In terms of training strategy,this thesis adds a variable to balance the generator and discriminator and improve the stability of the network.Experiments on public datasets show that the proposed method achieves leading performance.Aiming at the problem that unsupervised 3D human pose estimation does not depend on 3D data sets at all and has low accuracy,this thesis proposes an unsupervised 3D human pose estimation algorithm based on generative adversarial networks.The network presents a simple structure and is easy to train.In particular,the design of the random projection layer allows the network to distinguish the real 2D pose from the projected 2D pose from multiple directions.This thesis also introduces two human body geometric constraints,bone length constraints and joint angle constraints,which can reduce the fuzzy judgment caused by previous projections and make the 3D pose more accurate.Experiments on public datasets show that the proposed method achieves competitive results.