轻量化网络对人体行为识别模型的影响  

Influence of lightweight networks on human action recognition models

作　　者：刘雨辰 董建设[1] 蒙佳 梁成旭 LIU Yuchen;DONG Jianshe;MENG Jia;LIANG Chengxu(School of Information Technology Engineering,Tianjin University of Technology and Education,Tianjin 300222,China)

机构地区：[1]天津职业技术师范大学信息技术工程学院,天津300222

出　　处：《天津职业技术师范大学学报》2024年第4期25-30,共6页Journal of Tianjin University of Technology and Education

基　　金：天津市科技计划项目(23YDTPJC00320);天津市高等学校科技发展基金计划项目(2022ZD010);天津市信息传感与智能控制重点实验室基金项目(2023KFJJ04)。

摘　　要：针对人体行为识别研究中模型参数较多、复杂度较高的问题,提出在基于骨架的人体行为识别研究中,用几种轻量化的网络对时间图卷积进行改进。分别将瓶颈结构、可分离卷积、倒残差、沙漏块、involution算子、GhostNet结构引入时间图卷积,在尽可能保持模型识别精度的基础上,减少模型的参数。通过实验将提出的几种方法在公开数据集NTU RGB-D上与基准模型进行对比,结果表明:部分算法(瓶颈结构、可分离卷积、沙漏块、GhostNet结构)能够有效减少模型的参数,具有一定的可行性;可分离卷积在基于图卷积网络的人体行为识别模型中具有较好的效果。This paper addresses the challenges arising from the large number of model parameters and enormous complexity in human action recognition research.Several lightweight networks are used to improve Temporal Graph Convolutional Network(TGCN) in the research of skeleton-based human action recognition.The bottleneck structure,separable convolution,inverted residuals,hourglass blocks,involution algorithm,and GhostNet architecture are introduced into TGCN,respectively,aiming to reduce the number of parameters of the model while maintaining the recognition accuracy as much as possible.Through experiments,several methods proposed in this paper are compared with the baseline model on the public data set NTU RGB-D.The results indicate that several algorithms,notably the bottleneck structure,separable convolution,hourglass blocks,and GhostNet architecture,can effectively reduce the number of model parameters and demonstrate certain feasibility,and that separable convolution stands out for its superior efficacy in graph convolution-based human action recognition models.

关 键 词：行为识别 深度学习 轻量化 图卷积 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]