融合时空图卷积网络与非自回归模型的三维人体运动预测

当前人体运动预测的方法大多采用基于图卷积网络的自回归模型，没有充分考虑关节间的特有关系和自回归网络性能的限制，从而产生平均姿态和误差累积等问题。为解决以上问题，提出融合时空图卷积网络和非自回归的模型对人体运动进行预测。一方面利用时空图卷积的网络提取人体运动序列的局部特征，可以有效减少三维人体运动预测场景中的平均姿态问题和过度堆叠图卷积层引起的过平滑问题的发生；另一方面将非自回归模型与时空图卷积网络进行结合，减少误差累计问题的发生。利用Human3.6M的数据集进行80 ms、160 ms、320 ms和400 ms的人体运动预测实验。结果表明，NAS-GCN模型与现有方法相比，能预测出更精确的结果。

Three-dimensional human motion prediction combining spatiotemporal graph convolutional networks and non-autoregressive models

The current methods for predicting human motion mostly use autoregressive models based on graph convolutional networks, without fully considering the unique relationships between joints and the limitations of autoregressive network performance, resulting in issues such as average posture and error accumulation. To address the above issues, this paper proposed a fusion of spatiotemporal graph convolutional networks and non autoregressive models for predicting human motion. On the one hand, using a network of spatiotemporal graph convolutions to extract local features of human motion sequences can effectively reduce the occurrence of average pose problems and oversmooth problems caused by excessive stacking of graph convolutions in 3D human motion prediction scenes. On the other hand, it combined non-autoregressive models with spatiotemporal graph convolutional networks to reduce the occurrence of error accumulation problems. Conduct human motion prediction experiments using a Human3.6M dataset for 80 ms, 160 ms, 320 ms, and 400 ms. The experimental results indicate that the NAS-GCN model predicts more accurate results compared to existing methods.