华侨大学学报（自然科学版）2025，Vol.46Issue(3)：308-318,11.DOI:10.11830/ISSN.1000-5013.202409023
基于改进图卷积网络和人体骨架的扶梯场景危险行为识别
Dangerous Behavior Recognition in Escalator Scene Based on Improved Graph Convolutional Network and Human Skeleton
摘要
Abstract
To address the occlusion problem in the narrow environment of escalators and the accurate recogni-tion of similar human skeleton actions,a novel method called attention-guided multi-scale hierarchical edge ag-gregation sequential graph convolutional network(AMHGCN)is proposed to enable the graph neural network to capture missing human skeleton information from adjacent frames.Firstly,multi-scale features with differ-ent dilation rates are added to the temporal convolutional network,the extended seven branches can enhance the network's ability to extract features in the time domain.Secondly,hierarchical edge convolution is added after the multi-scale feature temporal convolutional network to expand local features to global features.Final-ly,a spatial channel attention mechanism is incorporated into each spatiotemporal graph convolutional block to strengthen the network's processing of spatial and channel information,making AMHGCN pays more atten-tion to the detailed features of different behaviors in the classification process and improves the classification accuracy.The evaluation of AMHGCN is conducted on the NTU RGB+D dataset and the escalator dangerous behavior dataset.The results show that compared to the baseline method STGCN++,AMHGCN achieves a significant improment in recognition accuracy on both the NTU RGB+D dataset and the escalator dangerous behavior dataset.关键词
扶梯/图神经网络/危险行为识别/多尺度特征/层次边缘卷积Key words
escalator/graph neural network/dangerous behavior identification/multi-scale feature/hierarchi-cal edge convolution分类
计算机与自动化引用本文复制引用
何建海,郑力新,臧佳明,庄琼云,潘书万..基于改进图卷积网络和人体骨架的扶梯场景危险行为识别[J].华侨大学学报（自然科学版）,2025,46(3):308-318,11.基金项目
福建省科技计划重点项目(2020Y0039) （2020Y0039）
黎明职业大学2022年度校级一般课题(自然科学类)(LZ202211) （自然科学类）
