[1]束 阳,李汪根,高 坤,等.基于轻量级语义信息融合的动作识别方法[J].计算机技术与发展,2023,33(06):181-188.[doi:10. 3969 / j. issn. 1673-629X. 2023. 06. 027]
　SHU Yang,LI Wang-gen,GAO Kun,et al.Action Recognition Method Based on Lightweight Semantic Information Fusion[J].,2023,33(06):181-188.[doi:10. 3969 / j. issn. 1673-629X. 2023. 06. 027]

点击复制

基于轻量级语义信息融合的动作识别方法()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年06期

页码:

181-188

栏目:

人工智能

出版日期:

2023-06-10

文章信息/Info

Title:

Action Recognition Method Based on Lightweight Semantic Information Fusion

文章编号:

1673-629X(2023)06-0181-08

作者:

束 阳; 李汪根; 高 坤; 王志格; 葛英奎

安徽师范大学 计算机与信息学院,安徽 芜湖 241002

Author(s):

SHU Yang; LI Wang-gen; GAO Kun; WANG Zhi-ge; GE Ying-kui

School of Computer & Information,Anhui Normal University,Wuhu 241002,China

关键词:

语义信息; 动作识别; 轻量级; 自注意力; 分流网络

Keywords:

semantic information; action recognition; lightweight; self-attention; distribution network

分类号:

TP391

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 06. 027

摘要:

针对目前大多数的动作识别方法使用深层网络训练模型导致模型参数量大、验证成本高以及语义信息利用不足等问题,提出一种基于轻量级语义信息融合的动作识别方法( LSIF-GCN) ,实现

了模型的轻量化和对语义信息的充分利用。 首先,LSIF-GCN 将数据预处理后的关节流、速度流和骨骼流三种不同的输入信息编码至高维空间后,经过一层图卷积操作,以达到特征增强和降低

维度的目的,再把三种信息流在通道维度上进行拼接融合。 然后,为了充分利用语义信息提取不同关节之间潜在的权重关系,提出一种“瓶颈型” 的四层图卷积模块。 最后,采用分流网络设计的

时间卷积模块,并引入自注意力机制,在减少模型参数量的同时也提高了网络的性能。 该模型具有简单的结构和训练过程,便于在低成本的嵌入式设备的实时动作识别系统中部署。 在 NTU-RGB+D 60 和 NTU-RGB+D 120 数据集上的大量实验表明,该方法不仅在识别精度和模型复杂度( 参数量和 GFLOPs) 上优于目前一些主流的轻量级方法,而且与一些近几年的 SOTA 方法相比

也具有一定的优势。

Abstract:

Aiming at the problems that most of the current action recognition methods use deep networks to train models,which leads tolarge amount of model parameters,high verification cost?

and insufficient utilization of semantic information,an action recognition methodbased on lightweight semantic information fusion ( LSIF-GCN) is proposed,which realizes the lightweight?

of the model and the full useof semantic information. First of all,LSIF-GCN encodes three different input information of joint flow,velocity flow and bone flow afterdata pretreatment into a?

high-dimensional space,and then goes through a layer of graph convolution operation to achieve the purpose offeature enhancement and dimension reduction. Then, the three information flows are spliced and fused in the channel dimension.Secondly,we propose a " bottleneck" four - layer graph convolution module to make full use of semantic information to extract the　potential weight relationship between different joints. Finally,the time convolution module designed by diversion network is adopted,andthe self-attention mechanism is introduced,which not only reduces the number of model parameters but also improves the network performance. The model has simple structure and training process,which can be easily deployed in real-time motion recognition system oflow cost embedded devices. A large number of experiments on the NTU - RGB + D 60 and NTU - RGB + D 120 dataset show that theproposed method not only outperforms some mainstream lightweight methods in recognition accuracy and model complexity ( parameternumber and GFLOPs) ,but also has certain advantages compared with some SOTA methods in recent years.

常用功能