基于GCN的局部增强人体骨骼行为识别算法

Locally enhanced human skeleton action recognition algorithm based on GCN

为使人体骨骼这种非欧几里得结构数据更好地应用于人体行为识别任务中,将多种模态的特征融合后作为模型的输入,探讨关节的全局与局部的信息共现性,构建一个完整的空间信息推理模型。通过使用多层时间膨胀卷积网络提取骨骼数据在时间维度中的上下文依赖关系,使模型可以兼具捕获短期和长期的时序上下文依赖关系,有效提高时间信息的提取能力。提出的分层次时空图卷积神经网络模型在NTU-RGB+D-60和NTU-RGB+D-120数据集上均取得了较好的识别准确率。

To better apply non-Euclidean structural data like human skeleton to human action recognition tasks,the features of multiple modes were taken as input to the model,global and local information co-occurrence of joints was explored,and a complete spatial information reasoning model was constructed.A multilayer temporal dilation convolution network was used to extract the context dependence of skeleton data in the temporal dimension,which enabled the model to capture both short-term and long-term time series context dependencies,and effectively improves the ability to extract temporal information.The hiera-rchical spatial-temporal graph convolution neural network model proposed achieves good recognition accuracy on NTU-RGB+D-60 and NTU-RGB+D-120 datasets.