Abstract:

Bone-based action recognition is widely concerned because it is not affected by the physical characteristics of the human body, simply and clearly conveys important information about human behavior recognition. Traditional application skeleton modeling usually relies on the artificial setting of traversal rules, which results in limited expressive power and less influential. Therefore, based on the recent popular Spatio-Temporal Graph Convolutional Network（ST-GCN） model, a new partitioning strategy for dividing skeleton joint points is proposed. Compared with the original partition method, this strategy strengthens the relationship between the relative positions of the body, which is conducive to improving the temporal and spatial correlation of skeleton joint point information. At the same time, during the training process, different iterative learning rates are set to further improve the recognition accuracy. The recognition effect is compared with the existing methods on two large-scale data sets Kinetics and NTU-RGB+D, the experimental results show the effectiveness of the proposed method.

Key words: behavior recognition, joint points, Spatio-Temporal Graph Convolutional Network（ST-GCN）, partition strategy, learning rate

摘要：

基于骨骼的动作识别因不受人体物理特征的影响，简单清晰地传达了人体行为识别的重要信息而受到广泛关注。传统的应用程序骨架建模通常依赖遍历规则的人为设置而导致表达能力有限和推广困难。因此，在近年来热门的时空图卷积网络（ST-GCN）模型基础上提出了一种新的划分骨架关节点的分区策略。该策略相比于原始分区方法加强了身体相对位置之间的关系，从而有利于提高骨架关节点信息在时间和空间上的关联。与此同时，在训练过程中通过设置不同的迭代学习率以进一步提高识别精度。在两个不同性质的大规模数据集Kinetics和NTU-RGB+D上与现有方法进行识别效果的比较，实验结果表明了该方法的有效性。

关键词: 行为识别, 关节点, 时空图卷积网络（ST-GCN）, 分区策略, 学习率