摘要
通过可穿戴传感器识别人体行为受到了广泛关注。现有的方法忽略了个体行为数据之间的潜在关系,尤其不能处理类内差异和类间相似的问题。为了解决这一限制,本文提出了具有特征相似性和个人特点的双通道混合图卷积网络(GCN)。一个通道通过特征图收集相似的活动信息,另一个通道根据个人特征图挖掘个人习惯对人类活动的内在影响。考虑到不同数据分布的差异,引入自注意机制对双通道进行加权,并根据不同的输入数据自适应调整两种拓扑的重要性,以提高网络的泛化性能。为了评估所提出的模型的性能,在UCI-HAR和WISDM数据集上进行了实验验证。结果表明:HSP-GCN的性能优于对比神经网络,F1分别为98.4%和96.5%,与现有工作相比有显著提高。
Human activity recognition through wearable sensors has received extensive attention.The existing methods ignore the potential relationship between individual behavioral data,especially cannot deal with the problem of intra-class difference and interclass similarity.In order to solve this limitation,a dual-channel hybrid graph convolutional network(GCN)with feature similarity and personal characteristics is proposed.One channel gathers similar activity information through feature graph,and the other channel mines the inherent influence of individual habits on human activity according to personal characteristics graph.Considering the difference of different data distribution,the self-attention mechanism(SAM)is introduced to weight the dual channels,and the importance of the two topologies is adaptively adjusted according to the different input data to improve the generalization performance of the network.In order to evaluate the performance of the proposed model,experimental validation is performed on the UCI-HAR and WISDM datasets.The experimental results show that the performance of HSP-GCN is better than contrastive neural network and the F1 is 98.4%and 96.5%,respectively,which is a significantly improved compared to the existing work.
作者
商樊淇
李志新
郇战
陈瑛
王永松
梁久祯
SHANG Fanqi;LI Zhixin;HUAN Zhan;CHEN Ying;WANG Yongsong;LIANG Jiuzhen(School of Microelectronics and Control Engineering,Changzhou University,Changzhou 213000,China;School of Computer Science and Artificial Intelligence,Changzhou University,Changzhou 213000,China)
出处
《传感器与微系统》
北大核心
2025年第3期138-142,共5页
Transducer and Microsystem Technologies
基金
国家自然科学基金资助项目(62201093)
江苏省研究生科研与实践创新计划资助项目(KYCX23_3070)。
关键词
深度学习
人体行为识别
图卷积神经网络
deep learning
human activity recognition
graph convolutional neural network
