.png){kind=icon}
返回
An improved spatial temporal graph convolutional network for robust skeleton-based action recognition
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-022-03589-y.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Skeleton-based action recognition methods using complete human skeletons have achieved remarkable performance, but the performance of these methods could significantly deteriorate when critical joints or frames of the skeleton sequence are occluded or disrupted. However, the acquisition of incomplete and noisy human skeletons is inevitable in realistic environments. In order to strengthen the robustness of action recognition model, we propose an Improved Spatial Temporal Graph Convolutional Network (IST-GCN) model, including three modules, namely Multi-dimension Adaptive Graph Convolutional Network (Md-AGCN), Enhanced Attention Mechanism (EAM) and Multi-Scale Temporal Convolutional Network (MS-TCN). Specifically, the Md-AGCN module can first adaptively adjust the graph structure according to different layers and the spatial dimension, temporal dimension, and channel dimension of different action samples to establish corresponding connections for long-range joints with dependencies. Then, the EAM module can focus on important information based on spatial domain, temporal domain and channel to further strengthen the dependencies between important joints. Finally, the MS-TCN module is used to enlarge the receptive field to extract more latent temporal dependencies. The comprehensive experiments on NTU-RGB+D and NTU-RGB+D 120 datasets demonstrate that our approach possesses outstanding performance in terms of both accuracy and robustness when skeleton samples are incomplete and noisy compared with the state-of-the-art (SOTA) approach. Moreover, the parameters and computational complexity of our model are far less than those of the existing approaches.
Keyword:
Action recognition
Adaptive graph
Multi-scale
Occlusion and noise
期刊
IF:
3.5
论文数:
7.6K
被引数:
1.7W
机构
引用论文
Predictively encoded graph convolutional network for noise-robust skeleton-based action recognition基于预测编码的图卷积网络，用于噪声鲁棒的基于骨架的动作识别.png)
APPLIED INTELLIGENCE
OAAI
APPLIED INTELLIGENCEIF3.5
OAAI分享收藏Evaluation of the Effect of Systolic Blood Pressure and Pulse Pressure on Cognitive Function: The Women's Health and Aging Study IIPLoS ONEOAAI
PLoS ONEIF0
OAAI分享收藏分享收藏
查看更多内容