【摘要】 人体行为识别是智能监控、人机交互等诸多应用领域的一项基本技术。人体骨骼的动态变化为人体行为识别提供了重要的信息。传统方法通常只是采取人工信息标注或遍历规则,从而导致模型的表征能力有限、泛化性能差。采用一种引入了残差项的动态骨架模型——基于残差连接的时空图卷积网络,不仅克服了以往方法的限制,而且能够学习骨骼数据中的时空模型。在大型骨骼NTU-RGB+D数据集上,该网络模型不仅提高了人体行为特征的表征能力,而且增强了泛化能力,取得了比现有的模型更好的识别效果。更多还原

【Abstract】 Human behavior recognition is a basic technology in many application fields such as intelligent monitoring and human-computer interaction. Dynamic changes of human skeleton provide important information for human behavior recognition. Traditional methods usually only adopt manual information annotation or traverse rules, resulting in limited representation capability and poor generalization performance of the model. This paper uses a kind of dynamic skeleton model with residual terms, which is a spatio-temporal graph convolution network based on residual connection. It not only overcame the limitations of previous methods, but also could learn the spatio-temporal model in the skeleton data. On the large skeleton NTU-RGB+D dataset, the network model improves the representation ability of human behavior characteristics, enhances the generalization ability and achieves better recognition results than the existing models.更多还原