Abstract: Dynamic multiscale graph neural networks have high motion prediction errors due to the low correlation between the internal joints of body parts and the limited perceptual fields. A multiscale global adaptive attention graph neural network for human motion prediction is proposed to reduce motion prediction errors. Firstly, a multi-distance partitioning strategy for dividing skeleton joint is proposed to improve the degree of temporal and spatial correlation of body joint information. Secondly, a global adaptive attention spatial temporal graph convolutional network is designed to dynamically enhance the network??s attention to the spatial temporal joints contributing to a motion in combination with global adaptive attention. Finally, this paper integrates the above two improvements into the graph convolutional neural network gate recurrent unit to enhance the state propagation performance of the decoding network and reduce prediction errors. Experimental results show that the prediction error of the proposed method is decreased on Human 3.6M dataset, CMU Mocap dataset and 3DPW dataset compared with state-of-the-art methods.

Key words: motion prediction, multi-distance partitioning strategy, global adaptive attention, spatial temporal graph convolution neural network, gated recurrent unit

摘要： 针对动态多尺度图神经网络的编解码网络中存在的身体部位内部关节点间关联度不高和感受野受限制导致运动预测误差偏高的问题，提出了一种用于人体运动预测的多尺度全局自适应注意力图神经网络，降低运动预测误差。提出了一种划分骨架关节点的多距离分区策略，用于提高身体部位关节点信息在时间和空间上的关联程度；提出了全局自适应注意力时空卷积神经网络，以动态地加强网络对某一动作有贡献的时空关节点的关注度；将上述两处改进集成到图卷积神经网络门控循环单元中，以增强解码网络的状态传播性能，并降低预测误差。实验表明，与最新方法相比，该方法在Human 3.6M、CMU Mocap和3DPW数据集上的预测误差都有所下降。

关键词: 运动预测, 多距离分区策略, 全局自适应注意力, 时空图卷积神经网络, 门控循环单元