基于ST-GCN的轻量级行为识别方法
Lightweight Behavior Recognition Method Based on ST-GCN
刘雪 1陈亚军 1吴玉娟1
作者信息
- 1. 西华师范大学电子信息工程学院 南充 637000
- 折叠
摘要
目前主流的一些行为识别方法存在着模型参数量大,计算复杂度过于复杂等问题,对于精确度与时效性很难同时满足.论文在时空图卷积网络的基础上:1)在输入中加入一个自训练注意力结构图学习节点隐性连接特征;2)空间和时间模块中分别加入脖颈结构和残差结构减少模型参数量与训练难度;3)空间与时间模块中加入通道挤压与激励机制学习全局特征提升识别率.在行为识别数据集NTU-RGB+D 60上的大量实验表明:模型计算复杂度和参数量减少4.7倍与4.4倍的同时,在CS与CV评价指标上精确率分别提升2.8%与3.1%.
Abstract
At present,some mainstream behavior recognition methods have some problems,such as large number of model pa-rameters and too complex computation,which are difficult to satisfy both accuracy and timeliness.The paper,based on spatiotempo-ral graph convolutional networks,introduces a self-training attention mechanism to learn the implicit connections of node features in the input,incorporates neck structures and residual structures in the spatial and temporal modules to reduce model parameters and training difficulty,adds channel squeeze and excitation mechanisms in the spatial and temporal modules to learn global features and improve recognition accuracy.A large number of experiments on the behavior recognition dataset NTU-RGB+D 60 show that the accuracy of CS and CV is increased by 2.8%and 3.1%,respectively,while the computational complexity and the number of parame-ters are reduced by 4.7 and 4.4 times.
关键词
轻量级/行为识别/骨骼/时空图卷积网络/通道挤压与激励机制Key words
lightweight/behavior recognition/bone/spatiotemporal graph convolutional network/channel extrusion and in-centive mechanism引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据