融合图卷积与Transformer的三维人体姿态估计网络

3D Human Pose Estimation Network Combining Graph Convolution and Transformer

两阶段的3D人体姿态估计方法因先进的2D姿态检测器而取得了显著进步,但深度信息的歧义性仍使这项任务极具挑战性。为解决此难题,提出了MGCNTrans的3D人体姿态估计网络。该方法采用2D-3D的提升策略。MGCNTrans网络融合了Transformer网络和空间通道图卷积网络的优势。该模型以多帧数据为输入,利用周围帧的信息来约束当前帧的姿态估计。在特征学习方面,图卷积网络被用于学习人体关节之间的物理连接关系,捕捉局部的空间特征。而Transformer网络则挖掘关节之间的隐式关系,提供全局的上下文信息。为提升模型性能,图卷积层融合了空间层和通道层,促使节点在局部和全局范围内更好地进行交互,增加特征多样性,并更准确地估计人体姿态。结果表明,MGCNTrans网络在3D人体姿态估计任务上取得了优越性能,证明了其有效性和先进性。

The two-stage 3D human pose estimation method has made significant progress due to advanced 2D pose detectors,but the ambiguity of depth information still makes this task extremely challenging.To solve this problem,a 3D human pose estimation network based on MGCNTrans is proposed.This method adopts a 2D-3D boosting strategy.The MGCNTrans network combines the advantages of Transformer network and spatial channel graph convolutional network.This model takes multiple frames of data as input and utilizes information from surrounding frames to constrain the pose estimation of the current frame.In terms of feature learning,graph convolutional networks are used to learn the physical connections between human joints and capture local spatial features.The Transformer network mines the implicit relationships between joints and provides global contextual information.To improve model performance,the graph convolutional layer integrates spatial and channel layers,enabling better interaction between nodes at both local and global scales,increasing feature diversity,and more accurately estimating human pose.The results show that MGCNTrans network has achieved superior performance in 3D human posture estimation task,which proves its effectiveness and progressiveness.